"When does it work ?" : An exploratory analysis of Transfer Learning for BCI

Transfer Learning is a critical topic of research in the BCI field. Its goal is to reuse data gathered in a previous session (source session) in order to reduce, or completely bypass, calibration in a new session (target session). Although many methods have been proposed to tackle this problem, little is known about what characteristics of the datasets should be taken into account in order to ensure good performance. In this paper, we perform an exploratory analysis to study the influence of some simple descriptors of the source and target datasets over the classification scores obtained with Transfer Learning. We observe that the discriminability of the data points in the target session plays an important role in determining how well the Transfer Learning will work, as opposed to that of the source session, which has no statistically significant role in most cases.