A Subject-Adaptive Brain State Decoding Model via Ensemble Transfer Learning

The cross-subject variability poses a great challenge to the practical application of the brain state decoding model. Although many transfer learning methods have been used to solve this problem, most of them directly combine existing subjects into a mixed source domain, ignoring the differences among multiple existing subjects. It's hard to align the target subject's data with the mixed source domain. Thus, we aim to reduce the cross-subject variability among different subjects and make full use of the rich information from them. We propose an ensemble transfer learning (ETL) method based on transfer joint matching to construct a subject-adaptive decoding model in an ensemble fashion. ETL can reduce the differences between the pairs of subjects, as well as the differences among multiple existing subjects. We found that many-to-one scheme could improve the performance with more data from multiple existing subjects, compared with one-to-one scheme, while the standard deviations of one-to-one schemes were much smaller. Moreover, the results of comparison methods and ablation experiments proved the effectiveness of our ETL method to decode brain state.