Electroencephalography (EEG) has been widely used for decoding user's mental states with the hope for practical real-world brain machine interfaces (BMIs). In typical BMIs, both spatial filters and classifiers must be learned from data in advance. A major obstacle in applications of BMIs into everyday-life contexts is the high variability of EEG signals between subjects or between recording sessions. The learning thus needs to be conducted separately for each user, even with an additional calibration step before every session. However, collecting enough amount of data in each user or in each session to achieve accurate decoding is so quite time-consuming that it imposes a large burden on BMI users. Reducing the high costs of data collection and also dealing with the inter-user or inter-session variabilities are thus crucial issues for developments of user-friendly BMIs.
In this study, we propose a novel method for decoding the mental state of a newly seen (target) subject without using the target subject's brain activities during task execution. This was achieved by effectively transferring the information from previously seen (training) subjects to the target subject through a computational technique to reduce the variabilities across subjects. The technique extracts from training subject the spatial bases common across subjects and sessions under the following two hypotheses: 1) brain activities can be expressed as a linear combination of a small number of (sparse) subject-session-independent spatial bases, 2) EEG signals derived from the brain activities are deformed by subject-session-dependent factors. When applied to a selective spatial attention task, our method successfully reduced the variabilities across subjects and sessions, so to achieve a decoding accuracy significantly better than the chance level. We also found that the extracted task-relevant spatial bases were reasonable on the basis of neuroscientific studies about the spatial attention.