Lowering Data Diversity can Accelerate Training: Case Studies in Synthetic Tasks
Keywords: synthetic tasks, data diversity, curriculum learning, data filtering, learning plateaus, batch gradients
TL;DR: Reducing data diversity can speed up training in a variety of synthetic settings.
Abstract: We identify a loss plateau at the start of training in the three synthetic settings of in-context linear regression, sparse parity, and fact memorization. While careful tweaks to the optimization algorithm can mitigate these plateaus, we find that a simpler orthogonal approach of *lowering the data diversity*, and in doing so, biasing the training distribution *away* from the test distribution, counter-intuitively also speeds up training. This connection between data diversity and training speed holds for three different diversity-*reducing* interventions across our varied synthetic settings. Our findings offer a new perspective on data filtering and curriculum learning for training machine learning models.
Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation 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: 12767
Loading