Dataset Condensation with Sharpness-Aware Trajectory Matching
Keywords: dataset condensation, meta-learning
Abstract: Dataset condensation aims to synthesise datasets with a few representative samples that can effectively represent the original datasets. This enables efficient training and produces models with performance close to those trained on the original sets. Most existing dataset condensation methods conduct dataset learning under the bilevel (inner and outer loop) based optimisation. However, due to its notoriously complicated loss landscape and expensive time-space complexity, the preceding methods either develop advanced training protocols so that the learned datasets generalise to unseen tasks or reduce the inner loop learning cost increasing proportionally to the unrolling steps. This phenomenon deteriorates when the datasets are learned via matching the trajectories of networks trained on the real and synthetic datasets with a long horizon inner loop. To address these issues, we introduce Sharpness-Aware Trajectory Matching (SATM), which enhances the generalisation capability of learned synthetic datasets by minimising sharpness in the outer loop of bilevel optimisation. Moreover, our approach is coupled with an efficient hypergradient approximation that is mathematically well-supported and straightforward to implement along with controllable computational overhead. Empirical evaluations of SATM demonstrate its effectiveness across various applications, including standard in-domain benchmarks and out-of-domain settings. Moreover, its easy-to-implement properties afford flexibility, allowing it to integrate with other advanced sharpness-aware minimisers. We will release our code on GitHub.
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.
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: 4382
Loading