IEEE Computational Intelligence Magazine - May 2018 - 73

C. Implementation

We adopted the implementations of the
algorithms as proposed for EM [25],
LLP [26] and the MIX method [27].
Please note the following. Since the EM
algorithm relies on a good (random)
initialization, Kindermans et al. [25] proposed to initialize five classifier pairs in
parallel, thus, increasing the chance of
having a good random initialization.
They also proposed to optimize the perclass variance b directly. In contrast, the
goal of the study by Verhoeven et al. [27]
was to compare the mean estimations of
the three methods. Hence, the authors
of the latter study used only one classifier and applied the same linear classifier with covariance-shrinkage by
Ledoit-Wolf [58], [71] to all three classifiers. In this paper, we used the original implementation of the EM
algorithm with five randomly initialized pairs of classifiers because we observed a better performance and hence,
have a fairer comparison.
For the online exper iment, the
algorithms were implemented in the
BBCI toolbox [40] in Matlab. A simplified code version for offline analysis
(without external toolboxes) is available
for Matlab2.
D. Experimental Setup, Data
Acquisition and Processing

Subjects were sitting at 80 cm distance
from a 24-inch flat screen. The EEG signals from 31 passive Ag/AgCl electrodes
(EasyCap) were recorded, which were
placed according to the extended 10-20
system, and whose impedances were
kept below 20 kX. The signals were
recorded and amplified by a multichannel EEG amplifier (BrainAmp DC,
Brain Products) at a sampling rate of
1 kHz. An optical sensor on the screen
indicated the exact starting time point
of each highlighting event.
The collected data was band-pass filtered with a third order Chebyshev Type
II filter between 0.5 and 8 Hz and
downsampled to 100 Hz. Epochs were
windowed to [−200, 700] ms relative to
2
Github repository: https://github.com/DavidHueb
ner/Unsupervised-BCI-Matlab.

The MIX method performs best for almost all subjects
and is able to consistently reach a high decoding
accuracy with an average of around 80% after data
of around seven characters has been recorded.
the stimulus onset and corrected for
baseline shifts observed in the interval
[−200, 0] ms. For each channel, the mean
amplitudes of six intervals ([50, 120],
[121, 200], [201, 280], [281, 380], [381, 530]
and [531, 700] ms relative to the stimulus
onset) were computed as features. As
strict instructions of participants during the online session regarding the
avoidance of obvious artifacts seemed
effective, we refrained from rejecting any
epochs in the preprocessing of the
online session and offline analysis after
the experiment. Apart from the described
ones, no further preprocessing steps
were applied.
E. Performance Scores

The accuracies of the three different
unsupervised classifiers were assessed
with two different metrics. First, the
spelling accuracy was computed, which
simply indicates whether a character was
spelled correctly or incorrectly. Second,
we computed the Area Under the
Curve (AUC) of the receiver operating
characteristic curve for discriminating
between target and non-target epochs.
The range of the AUC is between 0 and
1, where 0.5 indicates the theoretical
chance level and 1 indicates perfect separability, i.e., each event can be classified
correctly as target or non-target.
IV. Experimental Results

We start by presenting the results of the
online study. We observed that the
group-averaged visual ERP responses
upon target and non-target stimuli (data
not shown) are very similar to the ones
reported in our previous work [26] with
respect to the latencies of ERP peaks,
their amplitudes and spatial locations of
peaks on the scalp. This similarity is
expected, as we have used the same
highlighting scheme and a similar group
of subjects.

Regarding classification, Fig. 3A
shows the target vs. non-target classification accuracies for each subject and each
of the three unsupervised learning
method and Fig. 3B shows the grand
average over the 12 subjects. While LLP
reliably improves in the beginning but
only shows slow learning over time, the
EM algorithm performs more dichotomous. Depending on the random initialization, the classifier can either find
the projection very early (S7) or only
relatively late (S6, S9). The MIX method
performs best for almost all subjects and
is able to consistently reach a high
decoding accuracy with an average of
around 80% after data of around seven
characters has been recorded. We would
like to emphasize that seven characters
correspond to only 168 s of unsupervised training time or 476 unlabeled
epochs which suffice to reliably estimate
attended characters (see Fig. 3C). The
characteristic behaviors of the three classifiers also transfer to the spelling accuracy which is depicted in Fig. 3C.
Having found that the MIX method
is outperforming the two competing
unsupervised learning methods by a large
margin in the online study, the question
remains how well it competes with a
supervised classifier. We compared the
unsupervised MIX performance with
supervised shrinkage-regularized LDA
classifier [58] which is a highly competitive supervised classifier in the field of
BCIs [59]. As no supervised classifier was
used in the online experiments, we could
realize such a comparison only in a posthoc offline re-analysis of the data. Both
classifiers were trained on the first N - 1
characters and tested on the Nth character. Fig. 4 shows the results.
We tested the null hypothesis that
both single epoch classification accuracies come from the same distribution.
The non-parametric Wilcoxon rank sum

may 2018 | IEEE ComputatIonal IntEllIgEnCE magazInE

https://www.github.com/DavidHuebner/Unsupervised-BCI-Matlab https://www.github.com/DavidHuebner/Unsupervised-BCI-Matlab

Table of Contents for the Digital Edition of IEEE Computational Intelligence Magazine - May 2018