Gradient-based Optimization of Dataset Mixtures
Keywords: dataset selection, data valuation, foundation models
Abstract: Modern state-of-the-art machine learning models are often trained using a combination of heterogeneous data sources. However, the utility of different data sources as support for learning some target tasks is often not equivalent, motivating the need for automated methods of optimizing the relative contribution of each data source to the model. In this work, we propose a dataset optimization strategy that slices a normal model training step into a series of data source-specific updates and splices them back together in an optimal manner with respect to the loss on some target task dataset. We demonstrate the effectiveness of our algorithm across different scenarios and domains, including classification problems for vision models and for next-token prediction tasks in the language domain.
Primary Area: transfer learning, meta learning, and lifelong learning
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/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
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: 11577
Loading