GenDataAgent: On-the-fly Dataset Augmentation with Synthetic Data

ICLR 2025 Conference Submission312 Authors

Keywords: supervised learning, classification, computer vision, synthetic data, generative AI, responsible AI, fairness

TL;DR: GenDataAgent is an on-the-fly framework for augmenting image classification datasets, which emphasize the creation of focused, diverse, and in-distribution synthetic data.

Abstract:

Synthetic data is increasingly employed for training dataset augmentation in computer vision. However, prior works typically perform a uniform search across the entire category space, overlooking the interaction between synthetic data generation and downstream task training. Furthermore, balancing the diversity of synthetic data while ensuring it remains within the same distribution as real data (i.e., avoiding outliers) remains a significant challenge.
In this work, we propose a generative agent to augment target training datasets with synthetic data for model fine-tuning. Our agent iteratively generates relevant data on-the-fly, aligning with the target training dataset distribution. It prioritizes sampling diverse synthetic data that complements marginal training samples, with a focus on synthetic data that exhibit higher variance in gradient updates. Evaluations across diverse supervised image classification tasks demonstrate the effectiveness of our approach.

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: 312