Information Technology Reference
In-Depth Information
Table 1.1
The experiment was divided into six blocks of trials. Blocks were named after the
mental task the subjects were instructed to perform
Block
Subject 1
Subject 2
Subject 3
1
Active
Passive
Counting
2
Passive
Counting
Active
3
Counting
Active
Passive
4
Active
Passive
Counting
5
Passive
Counting
Active
6
Counting
Active
Passive
cation task was to determine the class of 2-second multi-channel
EEG segments, where class (1) = active listening, class (2) = passive listening and
class (3) = counting task.
The counting task was included as a control task to determine whether the EEG
features that might allow for the differentiation between the imagery and relaxed
listening tasks are not merely a function of a concentrating versus a non-concen-
trating state of mind.
Only the last four seconds (i.e. the second half of each trial) were considered for
analysis. These 4-second long segments were further divided into two 2-second
long segments. Thus, each trial yielded two segments. There were 120 trials for
each of the three conditions and each subject produced a total of 720 segments: 240
segments for each condition. The data are randomly partitioned into training set and
testing set with split ratio of 9:1, resulting in 648 training segments and 72 testing
segments.
We employed a linear auto-regression algorithm to represent the EEG data in a
compressed form in terms of estimations of spectral density in time (Anderson and
Sijercic
1996
; Peters et al.
1997
). Then, a classic single hidden-layer static neural
network (multi-layer perceptron), with variable number of hidden units and up to
three output units, was used for the classi
The classi
cation task. The network was trained in
batch mode for 50 epochs, using a scaled conjugate gradient algorithm, as described
by Bishop (
1995
). The data were divided into two sets: a training set
E
and a test set
T
. The training set
E
was used to train the neural network to recognise the mental
tasks of the elements that were left in
T
. In total, there were 768 inputs to the
network. The network was reset, retrained and reassessed 10 times with different
permutations of training and testing segments.
Classi
cations were made between 2-second long multi-channel segments
belonging to pairs of conditions (for 2-way classi
cation) and to all three conditions
(for 3-way classi
cation). The average classi
cation scores, including con
dence
limits and standard deviation, for each subject are shown in Table
1.2
.
Remarkably, the classi
cation scores are above 90 % accuracy. We acknowledge
that these results may not sound statistically robust because the experiment
involved only three subjects. Nevertheless, they encouraged me to work towards the
implementation of a BCMI on the assumption that the system would be capable to
