Least Disagree Metric-based Active Learning

Seong Jin Cho, Gwangsu Kim, Junghyun Lee, Chang D. Yoo

Keywords: active learning, uncertainty, disagree metric, diversity

Abstract: The most popular class of active learners today queries for the labels of the samples for which the prediction is most uncertain and uses the labeled samples to update its prediction. Unfortunately, quantifying uncertainty is an open question. This paper mathematically defines uncertainty in terms of the least disagree metric (LDM), which is the smallest perturbation required to alter the sample prediction. Based on this metric, the predictor is updated by querying the label of the most uncertain samples. Given a finite-sized training set, empirical LDM is incorporated into an active learning algorithm and used to approximate the theoretical LDM of each sample. Theoretical convergence properties between the empirical and the mathematical definition of LDM are provided. Experimental results show that our algorithm mostly outperforms other high-performing active learning algorithms and leads to state-of-the-art performance on various datasets and deep networks.

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.

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: Yes

Please Choose The Closest Area That Your Submission Falls Into: General Machine Learning (ie none of the above)

TL;DR: The uncertainty-based active learning algorithm based on the least disagree metric, which is the smallest perturbation required to alter the sample prediction.