Rethinking Dataset Quantization: Efficient Core Set Selection via Semantically-Aware Data Augmentation
Keywords: Coreset Selection, Dataset Quantization, Data Augmentation, Efficient Deep Learning, Semantically-Aware Augmentation
TL;DR: This paper proposes an efficient core set selection method based on semantically-aware data augmentation.
Abstract: Dataset quantization (DQ) is an innovative coreset selection method to choose representative subsets from large-scale datasets, such as ImageNet. Although DQ has made significant progress, it heavily relies on large pre-trained models (like MAEs), leading to substantial additional computational overhead. We first identify that removing this pre-trained MAE model degrades DQ’s performance and increases the variance in model training. Where MAE plays a crucial role in introducing prior knowledge and implicit regularization into the training process. Second, we investigate a data augmentation scheme that can simulate the steps of pixel compression and reconstruction in DQ by simply using a randomly initialized ResNet model. This randomly initialized ResNet model can take advantage of the inductive bias of CNNs to locate the semantic object region and then replace the other region with other images. Therefore, we can use a random model or trained model in the early training stage to enhance semantic diversity while selecting important samples. We remove the module that contains the pre-trained MAE model and integrate the data augmentation scheme into the DQ pipeline, which formulates a new simple but efficient method, called DQ v2. Our method achieves performance improvements across multiple datasets, such as ImageNette, CUB-200, and Food-101.
Primary Area: applications to computer vision, audio, language, and other modalities
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: 12709
Loading