Аннотация
The efforts to reduce user-specific training and calibration time of a Brain-Computer Interface (BCI) have been ramped up in the last several years. Such efforts have led to several studies that aimed at constructing subject-independent classifiers of mental intent; that is to say, BCIs that require no user-specific calibration process. This is a challenging problem due to major inter-subject variability of brain signals. In this work, we are primarily concerned with the same scenario; that is, constructing accurate subject-independent classifiers of mental intent. Given successful applications of convolutional neural network (CNN) in various areas including BCI, the utility of CNN in our work is not surprising. At the same time, the potential applications of ensemble learning in pattern recognition is not new. However, what is critical in ensemble-deep learning is how the ensemble per se is formed. The computational complexity involved in ensemble-deep learning is justified if the strategy used in forming the ensemble leads to a noticeable better classification performance. In this work, we propose multi-subject ensemble CNN (MS-En-CNN) classifier for zero-calibration BCI. This is an ensemble of CNN base classifiers that are trained using data collected from multiple subsets of training subjects. Our empirical results based on P300 visual-evoked potentials show that constructing MS-En-CNN leads to a significantly better subject-independent classification performance with respect to the average performance of the base CNN classifiers. Not only that, it is also shown that MS-En-CNN considerably improves (∼ 4% in our application) the average classification accuracy with respect to a single CNN that is trained based on pooled data from training subjects.