| Presentation | 2021-03-04 ［招待講演］因果メカニズム転移による小標本ドメイン適応 , |
|---|---|
| PDF Download Page |  PDF download Page Link |
| Abstract(in Japanese) | (See Japanese page) |
| Abstract(in English) | In the machine learning tasks where the training data is scarce, domain adaptation (DA) is a promising methodology that exploits the samples of relevant but different probability distributions (the source domains) in addition to the sample of the distribution for which one wants to eventually make inferences (the target domain). To justify DA methods, some assumption on the relations between the source and the target distributions (the transfer assumption) has to be postulated. Therefore, the central question in considering DA methods is "under what transfer assumptions, is the DA possible?" Most of the existing DA methods assume parametric forms of the distribution shift, that some conditional distributions are identical, or that some discrepancy between the source distribution and the target distribution is small in some measure. However, these assumptions may preclude the possibility of knowledge transfer among intricately shifted distributions or apparently very different distributions. In this presentation, as a candidate transfer assumption that may alleviate such a limitation, I consider a meta-distributional transfer assumption. Concretely, we assume that the source domains and the target domains have a shared structural equation model (a statistical causal model) behind the data distributions, and we propose a DA method called the causal mechanism transfer (CMT) that can be used in this scenario. In CMT, we pose the transfer assumption on the data-generating processes, instead of the data distributions that result from them. As a result, the assumption enables statistically justified DA among apparently very different data distributions. In this presentation, I will mainly introduce our recent research on few-shot domain adapting regression, a DA problem setup where we are given a large number of labeled source domain data and few but labeled target domain data, where we proposed a causal mechanism transfer and showed its validity from both theoretical and experimental perspectives. |
| Keyword(in Japanese) | (See Japanese page) |
| Keyword(in English) | Domain Adaptation / Structural Causal Model / Nonlinear Independent Component Analysis |
| Paper # | IBISML2020-59 |
| Date of Issue | 2021-02-23 (IBISML) |
| Conference Information | |
| Committee | IBISML |
|---|---|
| Conference Date | 2021/3/2(3days) |
| Place (in Japanese) | (See Japanese page) |
| Place (in English) | Online |
| Topics (in Japanese) | (See Japanese page) |
| Topics (in English) | Organized and general sessions on machine learning |
| Chair | Ichiro Takeuchi(Nagoya Inst. of Tech.) |
| Vice Chair | Masashi Sugiyama(Univ. of Tokyo) / Koji Tsuda(Univ. of Tokyo) |
| Secretary | Masashi Sugiyama(AIST) / Koji Tsuda(NTT) |
| Assistant | Atsuyoshi Nakamura(Hokkaido Univ.) / Shigeyuki Oba(Miidas) |
| Paper Information | |
| Registration To | Technical Committee on Infomation-Based Induction Sciences and Machine Learning |
|---|---|
| Language | JPN-ONLY |
| Title (in Japanese) | (See Japanese page) |
| Sub Title (in Japanese) | (See Japanese page) |
| Title (in English) | |
| Sub Title (in English) | |
| Keyword(1) | Domain Adaptation |
| Keyword(2) | Structural Causal Model |
| Keyword(3) | Nonlinear Independent Component Analysis |
| 1st Author's Name | |
| 1st Author's Affiliation | () |
| Date | 2021-03-04 |
| Paper # | IBISML2020-59 |
| Volume (vol) | vol.120 |
| Number (no) | IBISML-395 |
| Page | pp.pp.78-78(IBISML), |
| #Pages | 1 |
| Date of Issue | 2021-02-23 (IBISML) |