Batch Mode Active Learning Algorithm Combining with Self-training for Multiclass Brain-computer Interfaces

Abstract

In this paper, an batch mode active learning algorithm combining with the beneflts of self-training for solving the multiclass Brain-computer Interface (BCI) problem, which initially only needs a small set of labeled samples to train classiflers. The algorithm applied active learning to select the most informative samples and self-training to select the most high confldence samples, respectively, according to the proposed novel uncertainty criterion and confldence criterion for boosting the performance of the classifler. Experiments on the Dataset 2a of the BCI Competition IV, which demonstrate our method achieves more improvement than Active Learning (AL) and Random Sampling (RS) when the same amount of human efiort is sacriflced.