Dataset Projection: Finding Target-aligned Subsets of Auxiliary Data

Eric Wong, Kai Yuanqing Xiao, Aleksander Madry

Keywords: datasets, auxiliary data, dataset projection

TL;DR: We project datasets to find subsets of auxiliary datasets that are most aligned with a target dataset.

Abstract: To obtain more training data for a target task, one can draw upon related but distinct datasets, or auxiliary datasets. We put forth the problem of dataset projection---finding subsets of auxiliary datasets that are most aligned with a target dataset. These so-called projected datasets can be used as training data to improve performance on target tasks while being substantially smaller than the auxiliary dataset. We then develop a framework for solving such dataset projection problems and demonstrate in a variety of vision and language settings that the resulting projected datasets, when compared to the original auxiliary datasets, (1) are closer approximations of target datasets and (2) can be used to improve test performance or provide analysis for the target datasets.

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.

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: Yes

Please Choose The Closest Area That Your Submission Falls Into: General Machine Learning (ie none of the above)