Active Learning with Partial Labels

Fei Zhang, Junjie Ye, Lei Feng, Zhongwen Rao, Jieming Zhu, Marcus Kalander, Chen Gong, Jianye HAO, Bo Han

Keywords: weakly supervised learning, active learning, partial label learning

TL;DR: we propose a new problem setting named active learning with partial labels, where the oracle provides partial labels to the selected samples.

Abstract: In this paper, we for the first time study a new problem setting called active learning with partial labels (ALPL), where an oracle provides the query samples with a set of candidate labels that contains the true label. Such a setting relaxes the oracle from the demanding labeling process. To address ALPL, we firstly propose a firm and intuitive baseline by directly adapting a state-of-the-art method for learning with partial labels to train the predictor, which can be seamlessly incorporated into existing AL frameworks. Inspired by human inference in cognitive science, we propose to improve the baseline by exploiting and exploring counter examples (CEs) to relieve the overfitting caused by a few training samples in ALPL. Specifically, we propose to construct CEs by reversing the partial labels for each instance, learning from which we propose a simple but effective WorseNet. By leveraging the distribution gap between WorseNet and the predictor, both the predictor itself and the sample selection process can be improved. Experimental results on five real-world datasets and four benchmark datasets show that our proposed methods achieve comprehensive improvements over ten representative AL frameworks, highlighting the superiority and effectiveness of CEs and WorseNet.

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: Deep Learning and representational learning