An Efficient Query Strategy for Active Learning via Optimal Transport
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.
Keywords: Active Learning, Optimal Transport
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
TL;DR: We propose an efficient active query strategy based on optimal transport called AQOT and we empirically show it's a broad-spectrum active query strategy.
Abstract: Active Learning (AL) aims to reduce labeling costs by iteratively querying instances. Existing AL methods typically query instances based on either informativeness or representativeness. Only considering informativeness leads to sample bias. Only considering representativeness leads to query amount of instances before the optimal decision boundary is found. It is essential to consider both when querying instances. However, current hybrid methods are also time-consuming. To query instance efficiently while considering both informativeness and representativeness, we propose an efficient active query strategy based on optimal transport called Active Query by Optimal Transport (AQOT). Optimal Transport (OT) enables us to measure the difference between two distributions efficiently, allowing us considering the distribution of instances easily. Via entropy regularization, we can solve OT efficiently. Specifically, we make use of the sparseness of the solution of OT to querying the most informative instance while considering representativeness. Additionally, we introduce a dynamic adjustment to AQOT. By concatenating AQOT to multiple classification models, we show AQOT is a broad-spectrum active query strategy. Experimental results demonstrate that our method surpasses state-of-the-art active learning methods and shows high efficiency.
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: 7217
Loading