A Multi-Candidate Batch Mode Active Learning Approach

Xinyu Wang, Junli Wang, Chungang Yan

Abstract: Active learning (AL) is a popular method for addressing the problem of high labeling during sample training. Its main issue is determining how to establish a sample selection method. Existing research methodologies combine sample diversity and uncertainty, but they only study one candidate sample set, making it difficult to weigh the proportion of the two types of tactics in different circumstances. Therefore, this study proposes a batch-mode active learning method based on the multi-candidate-set that incorporates the screening findings of multiple pre-candidate sets while improving strategy stability. To efficiently screen out representative samples, a K-center screening strategy is adopted. On different image datasets, this method outperforms the baseline strategies and is more stable than the single candidate set strategy.