Meta Domain Reweighting for Partially Known Out-of-Distribution Generalization
Keywords: Out-of-Distribution Generalization, Domain Reweighting, Meta-Learning
TL;DR: This paper proposes a meta-learning-based reweighting method that automatically determines an effective reweighting of training samples to address known and unknown distribution shifts simultaneously.
Abstract: Distribution shift poses a significant challenge for modern machine learning methods when applied to real-world scenarios.
Existing researches typically assume that either unlabeled data from the target domain is available (domain adaptation) or nothing is known about the target dataset (domain generalization).
However, distribution shifts are often caused by environmental changes or human intervention, and some facets of the distribution shift can be predicted while others remain unknown.
To address this issue of partial knowledge in out-of-distribution generalization, this paper proposes a model-agnostic reweighting method named \emph{Meta Domain Reweighting for Partially Known Out-of-Distribution Generalization} (PKOOD).
Specifically, we utilize a bilevel meta-learning framework to simulate the known distribution shift and automatically determine an effective reweighting of the training samples to achieve strong generalization performance on unknown test datasets.
Additionally, we derive the upper bound of the risk gap between the reweighted training samples and the target dataset theoretically and incorporate it as a regularizer to guide loss design for reducing the variance and bias of both known and unknown distribution shifts.
The proposed method is evaluated on a real-world people income prediction dataset Adult and a recent out-of-distribution image classification benchmark NICO++, demonstrating its superiority over state-of-the-art algorithms regarding partially known out-of-distribution generalization performance.
Supplementary Material: pdf
Primary Area: causal reasoning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 1686
Loading