A Defense of One-Step Learning: Examining Single-Batch Distillations

ICLR 2025 Conference Submission7563 Authors

Keywords: distillation, interpretability, explainability, compression, cost surface, loss landscape

TL;DR: To understand dataset distillation, we examine distilled data instances and the cost surfaces produced by the distilled dataset.

Abstract: Dataset distillation produces a compressed synthetic dataset that approximates a large dataset or other learning task. A model can be trained on a distillation in a single gradient descent step. Conventional wisdom suggests that single-step learning is not generalizable and should yield poor performance; yet, distillation defies these expectations with good approximations of full direct-task training for a large distribution of models. In order to understand how distilled datasets can perform one-shot learning, we examine the distilled data instances and the cost surfaces produced by the distilled datasets. We demonstrate that the distilled dataset not only mimics features of the true dataset but also produces cost surfaces such that one-step training leads models from the initialization space into local minima of the true task's cost surface. This shows how one-step learning's counter-intuitive success is not only reasonable but also the expected outcome of dataset distillation.

Primary Area: interpretability and explainable AI

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: 7563