Adaptive Brain-Machine Interface of Brain-Controlled Vehicles Using Semi-MIM and TSVM

Brain-machine interfaces (BMIs) have been developed for healthy individuals to control external devices. However, like all the existing BMIs, a time-consuming training process is required. To address this problem, a semi-supervised decoding framework is proposed to develop an adaptive BMI. The adaptive BMI is firstly initialized using a small labeled training set, and then increasingly adjusts itself by updating with newly collected unlabeled electroencephalogram (EEG) samples. The semi-supervised decoding framework starts with a semi-supervised mutual information maximization (semi-MIM) method to select optimal features and then uses the transductive support vector machine (TSVM) for classification. Experimental results show that the proposed semi-supervised framework performs better than other semi-supervised approaches and enables the adaptive BMI to catch up with the performance of the supervised learning-based BMI. Since the adaptive BMI uses a smaller training set, it can significantly reduce the training effort.