Four eyes see more than two: Dataset Distillation with Mixture-of-Experts
Keywords: dataset distillation, mixture-of-experts
Abstract: The ever-growing size of datasets in deep learning presents a significant challenge in terms of training efficiency and computational cost. Dataset distillation (DD) has emerged as a promising approach to address this challenge by generating compact synthetic datasets that retain the essential information of the original data. However, existing DD methods often suffer from performance degradation when transferring distilled datasets across different network architectures (i.e. the model utilizing distilled dataset for further training is different from the one used in dataset distillation). To overcome this limitation, we propose a novel mixture-of-experts framework for dataset distillation. Our goal focuses on promoting diversity within the distilled dataset by distributing the distillation tasks to multiple expert models. Each expert specializes in distilling a distinct subset of the dataset, encouraging them to capture different aspects of the original data distribution. To further enhance diversity, we introduce a distance correlation minimization strategy to encourage the experts to learn distinct representations. Moreover, during the testing stage (where the distilled dataset is used for training a new model), the mixup-based fusion strategy is applied to better leverage the complementary information captured by each expert. Through extensive experiments, we demonstrate that our framework effectively mitigates the issue of cross-architecture performance degradation in dataset distillation, particularly in low-data regimes, leading to more efficient and versatile deep learning models while being trained upon the distilled dataset.
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: 3431
Loading