A bayesian model for exploiting application constraints to enable unsupervised training of a P300-based BCI

Abstract

This work introduces a novel classifier for a P300-based speller, which, contrary to common methods, can be trained entirely unsupervisedly using an Expectation Maximization approach, eliminating the need for costly dataset collection or tedious calibration sessions. We use publicly available datasets for validation of our method and show that our unsupervised classifier performs competitively with supervised state-of-the-art spellers. Finally, we demonstrate the added value of our method in different experimental settings which reflect realistic usage situations of increasing difficulty and which would be difficult or impossible to tackle with existing supervised or adaptive methods.

Figure 1. Plot showing the average P300 response versus the average background EEG.

Figure 2. The speller matrix used in this work.

Source: BCI Competition II dataset description.

Figure 3. The projection of the EEG into one dimension produces two Gaussians.

Figure A shows the histogram of the used EEG features projected into one dimension. Figure B shows two Gaussians fitted to this histogram. One Gaussian for the EEG containing the P300 response, one Gaussian for the data without P300 response. The vector that was used in the projection was trained unsupervisedly on the data.

Figure 4. Scatter plots showing the quality of the classifier.

Quality is measured in either AUC or characters predicted correctly versus the data log likelihood. The data used in this plot is created in the OFF-US experiment on subject B using 5 repetitions.

Figure 5. Bar graph showing the performance, measured in AUC, on the test.

The classifier OFF-US is trained unsupervisedly on the test set. The BOUND is trained supervisedly on the test set without regularization.

Figure 6. Classifier improvement trough adaptation.

The initial classifier was trained unsupervisedly on the train set with 5 repetitions. The classifier was adapted to the EEG by feeding it the EEG character by character and performing EM on the original training set combined with the new EEG.

Figure 7. Plots showing the performance obtained by 3 single online initializations on subject B, each using a different number of repetitions to predict a character.

The horizontal axis represents the number of characters processed. The vertical axis represents how many of these characters were predicted correctly. The dashed line shows us how many characters the online classifier has predicted correctly (starting with an initially untrained classifier). The solid line shows how many characters the current classifier can predict correctly if we re-test it on all of the previously processed characters. The dash-dot line represents the upper bound on the performance which equals the number of characters seen.