IEEE Computational Intelligence Magazine - May 2018 - 75

domain is about as valuable as labeled
data for the training of a classifier once a
sufficient amount of data is available. We
also found that the online performance observations of LLP, EM and MIX
match well with the three main results
reported in an earlier offline study by
Verhoeven et al. [27] regarding (a) the
single epoch classification, (b) spelling
accuracies and (c) the observed relative
performance differences between the
three methods. We conclude that no
severe overfitting has occurred in the
offline study by Verhoeven et al. [27] and
that these promising results are indeed
transferable to the online case.
The unsupervised approaches have
two key advantages over traditional
supervised methods. First, the unproductive calibration time required for training
a supervised classifier becomes superfluous with unsupervised learners. This
time can be utilized on the desired application right from the start, even though
the performance needs to ramp up over
a few trials. Subjects perceived the rampup phase as challenging because of the
confusing feedback. Many subjects were
blaming themselves rather than the computer for the incorrect feedback. We
want to stress that for some applications,
e.g., text spelling, the text generated during the ramp-up period can be corrected
post-hoc by simply reapplying the
improved classifier at a later time point
onto the initially collected data [25].
Another approach to mitigate the rampup effect of the MIX method is, of
course, to also incorporate transfer learning. Its benefit for unsupervised learning
approaches has been shown by Kindermans and colleagues [70].
The second big advantage of unsupervised classifiers is the ability to continuously learn from unlabeled data. It is
well-known that EEG signals do not
only change from calibration to online
sessions [72], but are also non-stationary
over longer sessions [72] due to changing human factors (fatigue, motivation
and learning) [2], [73] or non-human
factors (drying gel leading to changing
impedances, changed environmental
conditions). As supervised classifiers are
fixed after training, non-stationarities

can cause supervised methods to deteriorate over a session, while unsupervised methods have the chance to adapt to
changing data distributions and maintain
or even improve their classification accuracy over time [25].
The biggest limitation of the presented unsupervised learning approaches is
that-so far-they are mostly restricted
to ERP data and are not directly applicable to, e.g., motor imagery data. The reason is that they explicitly utilize the rich
structure introduced by the ERP paradigm. For instance, the EM algorithm
exploits that one latent variable-the
selected symbol-uniquely determines
all target and non-targets epochs of a
trial. The LLP approach requires a slight
modification of the stimulation paradigm
in order to create different class proportions. Here, one could consider the
option of switching back to a spelling
matrix without visual blanks to avoid
highlighting unnecessary symbols after
the ramp-up phase [26]. While this
might slightly change the ERP responses, the continued unsupervised adaptation should be able to adapt to these
changes. Re-visiting existing ERP-based
BCI applications, however, makes evident, that such class proportion differences might be available in some
applications already without changing
the interface, e.g., in applications which
implement a two-step selection procedure where the number of symbols differs in the first and second selection step.
In general, we think that future work
should go towards jointly adapting the
paradigm and classifier by considering
the user, interface and decoder as a holistic system. A first conceptual attempt has
been made by Mladenovic´ et al. [74].
The MIX method is the result of
combining two unsupervised learning
ideas with complementary strengths and
weaknesses [27]. By reviewing other
attempts, we hope to further foster the
combination of different ideas. In the
past, Kindermans et al. [70] already proposed a joint Bayesian framework utilizing a language model, dynamic stopping
and transfer learning. These add-on
techniques can also be combined with
the unsupervised MIX method if

increased spelling speed is required. Certainly, this set could be further extended,
for instance by exploiting error-related
potentials [19]. We think that coping
with the low SNR in BCI data requires
the aggregation of information from different temporal and neuronal sources
and a careful exploitation of the underlying data constraints.
VI. Conclusion

We reviewed different strategies to learn
from unlabeled data in ERP-based BCIs.
There is clear evidence that unsupervised
adaptation outperforms non-adaptive
supervised classifier. We also found conceptually different learning strategies
based on predicted labels, additional user
input such as error-related potentials or
based on the exploitation of the underlying ERP data constraints. As demonstrated with the MIX method, combining
these ideas can substantially improve
unsupervised learning approaches. An
online study with 12 healthy subjects
showed that the MIX method is currently by far the most promising unsupervised learning approach which can even
compete with a supervised state-of-theart method that has the same amount of
training data and full label information
available after a short ramp-up. If a slight
modification of the ERP paradigm is
accepted by BCI users, then unsupervised
learning methods can in practice completely replace supervised methods. This
opens the opportunity for true plug &
play systems and the ability to learn from
large unlabeled data sets to find common
patterns and improve transfer learning.
Acknowledgment

DH and MT thankfully acknowledge
the support by BrainLinks-BrainTools
Cluster of Excellence funded by the
German Research Foundation (DFG),
grant number EXC 1086. DH and MT
further acknowledge the bwHPC initiative, grant INST 39/963-1 FUGG. PJK
gratefully acknowledges funding from
the European Union's Horizon 2020
research and innovation program under
the Marie Sklodowska-Curie g rant
agreement NO 657679. TV thankfully
acknowledges financial support from the

may 2018 | IEEE ComputatIonal IntEllIgEnCE magazInE

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