Presentation | 2023-12-21 Classification Error Analysis under Covariate Shift between Non-absolutely Continuous Distributions through neighbor-transfer-exponent Mitsuhiro Fujikawa, Youhei Akimoto, Jun Sakuma, Kazuto Fukuchi, |
---|---|
PDF Download Page | PDF download Page Link |
Abstract(in Japanese) | (See Japanese page) |
Abstract(in English) | Transfer learning is considered successful when increasing the source sample size decreases the target sample size needed to achieve sufficient accuracy. However, existing analyses on the transfer learning under covariate shift show the success of transfer only if the target distribution is absolutely continuous with respect to the source distribution. Such a situation frequently happen in real world. For example, in the image recognition task of detecting scratches on industrial components, there may be a slight difference in the shape specifications of the existing (source) and new (target) factories. In this study, we provide a novel theoretical technique that can explain the success of transfer between non-absolute continuous distributions, by taking advantage of the smoothness of the regression function and the shifting of data points considering classification. Experimental results demonstrate that our upper bound better characterizes sample complexity than existing upper bounds under the context of non-absolutely continuous distributions. |
Keyword(in Japanese) | (See Japanese page) |
Keyword(in English) | transfer learning / covariate shift / excess error / non-absolutely continuous / transfer-exponent |
Paper # | IBISML2023-39 |
Date of Issue | 2023-12-13 (IBISML) |
Conference Information | |
Committee | IBISML |
---|---|
Conference Date | 2023/12/20(2days) |
Place (in Japanese) | (See Japanese page) |
Place (in English) | National Institute of Informatics |
Topics (in Japanese) | (See Japanese page) |
Topics (in English) | machine learning, etc. |
Chair | Masashi Sugiyama(Univ. of Tokyo) |
Vice Chair | Toshihiro Kamishima(AIST) / Koji Tsuda(Univ. of Tokyo) |
Secretary | Toshihiro Kamishima(NTT) / Koji Tsuda(Hokkaido Univ.) |
Assistant | Yoshinobu Kawahara(Osaka Univ.) / Taiji Suzuki(Univ.of Tokyo) |
Paper Information | |
Registration To | Technical Committee on Information-Based Induction Sciences and Machine Learning |
---|---|
Language | JPN |
Title (in Japanese) | (See Japanese page) |
Sub Title (in Japanese) | (See Japanese page) |
Title (in English) | Classification Error Analysis under Covariate Shift between Non-absolutely Continuous Distributions through neighbor-transfer-exponent |
Sub Title (in English) | |
Keyword(1) | transfer learning |
Keyword(2) | covariate shift |
Keyword(3) | excess error |
Keyword(4) | non-absolutely continuous |
Keyword(5) | transfer-exponent |
1st Author's Name | Mitsuhiro Fujikawa |
1st Author's Affiliation | University of Tsukuba(Univ. of Tsukuba) |
2nd Author's Name | Youhei Akimoto |
2nd Author's Affiliation | University of Tsukuba(Univ. of Tsukuba) |
3rd Author's Name | Jun Sakuma |
3rd Author's Affiliation | Tokyo Institute of Technology(Tokyo Inst. of Tech.) |
4th Author's Name | Kazuto Fukuchi |
4th Author's Affiliation | University of Tsukuba(Univ. of Tsukuba) |
Date | 2023-12-21 |
Paper # | IBISML2023-39 |
Volume (vol) | vol.123 |
Number (no) | IBISML-311 |
Page | pp.pp.58-65(IBISML), |
#Pages | 8 |
Date of Issue | 2023-12-13 (IBISML) |