Single shot active learning using pseudo annotators

作者：

Highlights：

• A single shot setting of active learning is addressed, where all the required samples should be chosen in a single shot.

• Pseudo annotators, which uniformly and randomly annotate queried samples, are introduced to impel standard active learning algorithms to explore.

• The exploratory behavior is further enhanced by selecting the most representative sample via minimizing nearest neighbor distance between unlabeled samples and queried samples.

• Excellent performance of the proposed method in comparison with state-of-the-art approaches is demonstrated.

摘要

•A single shot setting of active learning is addressed, where all the required samples should be chosen in a single shot.•Pseudo annotators, which uniformly and randomly annotate queried samples, are introduced to impel standard active learning algorithms to explore.•The exploratory behavior is further enhanced by selecting the most representative sample via minimizing nearest neighbor distance between unlabeled samples and queried samples.•Excellent performance of the proposed method in comparison with state-of-the-art approaches is demonstrated.

论文关键词：Active learning,Pseudo annotators,Random labeling,Single shot,Exploration and exploitation,Minimizing nearest neighbor distance

论文评审过程：Received 16 May 2018, Revised 4 December 2018, Accepted 18 December 2018, Available online 20 December 2018, Version of Record 26 December 2018.

论文官网地址：https://doi.org/10.1016/j.patcog.2018.12.027