Computational Intelligence - February 2016 - 27

Feature extraction
For feature extraction, recorded EEG data was first spatially filtered using a surface Laplacian setup [35]. We did not employ
more sophisticated methods for spatial filtering, such as CSP or
beamforming, in order to keep the spatial filtering setup dataindependent. For each subject, trial and electrode, frequency
bands of 2 Hz width, ranging from 7-29 Hz, were then
extracted using a discrete Fourier transform with a Hanning
window, computed over the length of the trial. Logbandpower
within the last seven seconds of each trial for each frequency
band then formed the (128 # 12)-dimensional feature vector.
Classification performance
Here we show the efficiency of the proposed algorithms by
examining the effect of multitask learning and FD on classification performance. For all algorithms, one subject was successively chosen as the test subject and all other subjects were then
used for training. Test-specific training data of between 10 and
100 trials per condition were then given to each algorithm, and
the remaining trials out of 300 were used for testing. Multitask
learning was done using Algorithm 1 and Algorithm 3 with a
cross-validated m. Note that for all tested algorithms the feature
space was the full 128 channels, each with 12 feature bands.
We looked at two control algorithms to compare with the
multitask learning approaches. The first was to consider ridge
regression, which regularizes the regression method only by
penalizing the magnitude of the resultant weight vector (see
(5)) and can be seen as using an uninformed prior for the distribution of weight vectors; the second was to consider a support vector machine (SVM) trained on the same feature space.
We further tested both control algorithms two ways: Once
with pooled data and once with only subject-specific data. For
the pooled condition, all data from the training subjects was
concatenated to the training trials from the test subject to form
a combined training set, on which the control algorithms were
run. For the subject-specific condition, only training data from
the test subject was used to train the control algorithms. All
controls were compared to the multitask approaches, where the
learned prior mean(s) and covariance(s) were used to regularize
the least-squares regression method.
The following list summarizes the algorithms:

Classification Accuracy

0.80
0.75
0.70
MT
MT_FD
SVM
RR
RR_FD

0.65
0.60
0.55
0.50

100

150

200

Subject Specific Training Trials
Figure 2 Mean and STD (shaded) for classification accuracy of MT
and pooled conditions across the ten subjects. The control algorithms
were trained on data pooled across training subjects, and are compared against classification using Algorithm 1 (MT, solid blue) and
Algorithm 3 (MT_FD, solid red). Displayed control algorithms are ridge
regression using the standard regression method (RR, dashed blue),
ridge regression using the FD regression method (RR_FD, dashed red)
and SVM (SVM, solid black). The FD formulation of the multitask
learning has comparable performance with few training trials to
pooled regression and both multitask algorithms manage to improve
more than the pooled controls given a larger number of training trials.

0.80
Classification Accuracy

Dataset
Ten healthy subjects participated in the study (two females,
25.6 ! 2.5 years old). One subject had already participated
twice in other motor imagery experiments while all others
were naïve to motor imagery and BCIs. EEG data was
recorded from 128 channels, placed according to the extended
10-20 system with electrode Cz as reference, and sampled at
500Hz. BrainAmp amplifiers (BrainProducts, Munich) with a
temporal analog high-pass filter with a time constant of 10s
were used for this purpose. A total of 150 trials per class (left/
right hand motor imagery) per subject were recorded in pseudorandomized order, with no feedback provided to the subjects
during the experiment.

0.75
0.70
MT
MT_FD
SVM
RR
RR_FD

0.65
0.60
0.55
0.50

100

150

200

Subject Specific Training Trials
Figure 3 Mean and STD (shaded) for classification accuracy of MT
and single-subject conditions across the ten subjects. Classification
values for the multitask algorithms are identical to those shown in
Figure 2. The control algorithms were trained on data exclusively
from the test subject, and are compared against classification using
Algorithm 1 (MT, solid blue) and Algorithm 3 (MT_FD, solid red). Displayed control algorithms are ridge regression using the standard
regression method (RR, dashed blue), ridge regression using the FD
regression method (RR_FD, dashed red) and SVM (SVM, solid black).
The multitask algorithm with FD regression estimation performs better on average regardless of the number of trials, though single-subject ridge regression with the FD regression method manages to
equal its performance at 200 training trials.

❏
❏
❏
❏
❏

MT_FD: multitask learning with Algorithm 3
MT: multitask learning with Algorithm 1
RR: standard ridge regression
RR_FD: ridge regression using the FD regression method
SVM: SVM with a linear kernel given the full 128 # 12
feature space

results
The results for the pooled sub-condition can be found in Figure 2 and the results for the single-subject sub-condition can

February 2016 | Ieee ComputatIonal IntellIgenCe magazIne

Table of Contents for the Digital Edition of Computational Intelligence - February 2016