Self-Informed Generative Active Learning
Keywords: Active Learning, Large Language Model, Synthetic Data, Reinforcement Learning
TL;DR: We propose the Self-Informed Generative Active Learning (SIGnAL) framework which actively generates and selects data instances for annotation and downstream model training.
Abstract: Active learning has been a cost-efficient approach to obtaining high-performance AI models with fewer selective annotations. In scenarios where the acquisition of original unlabeled data poses significant challenges, active learning harnessing synthesized data instances is more promising than traditional pool-based methods. In this paper, we propose the Self-Informed Generative Active Learning (SIGnAL) framework as an effective solution to actively generate and select data instances for annotation and downstream model training. In SIGnAL, we propose to guide the data generation based on a reinforcement learning policy, where the generator is self-informed by the reward to generate more informative instances. In addition, we introduce an acquisition function that measures both the informativeness and relevance of instances. Such acquisition function can be transformed to the reward seamlessly for generator optimization. Our experiments on the text classification task validate the effectiveness of our framework, especially when the original data scale is limited.
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: 12726
Loading