Principled algorithms for domain adaptation and generalization
Abstract/Contents
- Abstract
- Machine learning models are increasingly applied to datasets different from the training datasets. The performance of models often degrades when tested on unseen scenarios. Empirically, many algorithms have been used for domain adaptation and generalization, but few methods have been able to surpass empirical risk minimization consistently on common benchmarks. Theoretically, traditional learning theory offers limited insights for distributional shift problems. The main goal of this thesis is to bridge the gap between the theory and practice for domain shift problems, and to develop principled algorithms that have better robustness guarantees. We study three domain shift problems with increased supervised from the target domain. We first study domain generalization where no target data is available during training. We show that feature-matching algorithms generalize better when the distinguishing property of the signal feature is indeed conditional distributional invariance. Next, we study domain adaptation where unlabeled target data is available. We show that self-training helps when the target is more diverse than the source. Lastly, we study active online learning under domain shift. We show that uncertainty sampling leads to better query-regret tradeoff when there is hidden domain structure. In all three problems, the synergy of explicit bias from the algorithm and implicit bias from the domain shift structure contributes to successful transfer between domains.
Description
| Type of resource | text |
|---|---|
| Form | electronic resource; remote; computer; online resource |
| Extent | 1 online resource. |
| Place | California |
| Place | [Stanford, California] |
| Publisher | [Stanford University] |
| Copyright date | 2023; ©2023 |
| Publication date | 2023; 2023 |
| Issuance | monographic |
| Language | English |
Creators/Contributors
| Author | Chen, Yining, (Researcher in computer science) |
|---|---|
| Degree supervisor | Ma, Tengyu |
| Thesis advisor | Ma, Tengyu |
| Thesis advisor | Ermon, Stefano |
| Thesis advisor | Hashimoto, Tatsunori |
| Degree committee member | Ermon, Stefano |
| Degree committee member | Hashimoto, Tatsunori |
| Associated with | Stanford University, School of Engineering |
| Associated with | Stanford University, Department of Computer Science |
Subjects
| Genre | Theses |
|---|---|
| Genre | Text |
Bibliographic information
| Statement of responsibility | Yining Chen. |
|---|---|
| Note | Submitted to the Computer Science Department. |
| Thesis | Thesis Ph.D. Stanford University 2023. |
| Location | https://purl.stanford.edu/dm107nr6713 |
Access conditions
- Copyright
- © 2023 by Yining Chen
- License
- This work is licensed under a Creative Commons Attribution Non Commercial 3.0 Unported license (CC BY-NC).
Also listed in
Loading usage metrics...