Abstract
In Brain-computer Interface (BCI), the detection of activations is based on the experience gained through calibration or training sessions prior to actual use to build the classification model. This gives rise to several problems that include inter-session variability and time fading of accuracy after calibration. In this work, we investigate a new approach for brain-computer interface data that requires no prior training. The basic principle of this new class of unsupervised techniques is that the trial with true activation signal within each block has to be different from the rest of the trials within that block. Hence, a measure that is sensitive to this dissimilarity can be used to make a decision based on a single block without any prior training. The new approach is applied to experimental data for P300-based BCI for both normal and disabled subjects and compared to the classification results of the same data using the conventional processing techniques requiring prior calibration. Performance in different experiments assessed using classification block accuracy suggests that this approach can reach accuracies not very far from those obtained with training while maintaining robust performance in practice.