Information Technology Reference
In-Depth Information
Single-Voxel Based Voter Classifier (SVVC)
As shown in Figure 5, a set of 6-dimensional features vectors along with their
labels of a voxel is given to train a single-voxel classifier to form an optimal
discrimination hyperplane which means each single-voxel classifier after training
can predict the unknown type of test data. Because of a lot of noises in fMRI
data, however, the result predicted only by one voxel is unacceptable. This is only
the first stage. The second stage treat each single-voxel classifier as a voter, and
the final decision is made by calculating the weighted sum of the classification
of all single-voxel-classifiers. The weight is corresponding to the performance of
classifications of each classifier, the higher classification accuracy a single-voxel-
classifier obtained, the higher weight will be assigned to it.
The classifiers mentioned above are all SVM classifiers, and the LIBSVM
toolkit [13] based on MATLAB platform was used to train these classifiers. The
parameters of training are as follows: svm type = nu SVC, kernel type = RBF,
nu = 0.5, C and gamma were adjusted in given interval [-10, 10] to get optimized
performance.
The Leave-One-Session-Out method was adopted to evaluate the validity of
the classifiers. As there were 5 scan sessions for each participant, we randomly
chose 4 sessions data to train the classifier and used the data in the left session
to test the classifier to get one validation. There were totally 5 validations, the
final accuracy for each participant was the average of the accuracy values in
these 5 validations.
3R su s
We performed 2 kinds of classification tasks, one was step classification (1-step
tasks vs. 2-steps tasks), and the other was complexity classification (simple tasks
vs. complex tasks).
3.1
Results of MVC
The results of classification based on MVC for all subjects and the average ac-
curacy are shown in Figure 6(a), in which the step classifiers denoted in grey
columns, the highest accuracy reaches to 80.9% in individual participants, and
the average accuracy across 16 participants is 70.3% [SD = 0.04], which is sig-
nificantly higher than the random classification chance level of 50%. As to the
complexity classifiers denoted in dark columns, the highest accuracy is 68.7% in
individual participants, and the average accuracy across 16 participants is 63.3%
[SD = 0.03], which is significantly higher than the random classification chance
level of 50%. The results show that the classification in step is more effective
than the one in complexity.
3.2 Results of SVVC
The results of classification based on SVVC for all subjects and the average
accuracy are shown in Figure 6(b), in which the step classifiers denoted in grey
