ALBS: An Active Learning Framework Based on Syncretic Sample Selection Strategy

Abstract

Machine learning has achieved outstanding performance in many fields, but its success heavily relies on the large number of annotated training samples. However, for many professional fields, data annotation is not only tedious and time consuming, but also demanding specialty-oriented knowledge and skills, which are not easily accessible. To significantly reduce the cost of annotation, we propose a novel active learning framework called ALBS. ALBS uses the syncretic strategy which incorporates "most discriminative" and "most representative" to seek "worthy" samples from unlabeled dataset and update the model incrementally to enhance the performance continuously. We have evaluated our method on two different audio datasets, demonstrating that the syncretic strategy can makes the promotion of model model's performance more robust and faster than the other strategies, and subsampling the historical labeled dataset can reduce unnecessary computing costs and storage space.