Figure 1.
The baseline algorithm, in which only user-specific training samples are used, and new training samples are generated randomly from the
classes online.
Figure 2.
The TL algorithm, in which primary and auxiliary training samples are used together in determining the optimal
in the kNN classifier.
Figure 3.
The ACS algorithm, in which the classes from which new training samples are generated are determined based on per-class cross-validation performance.
Figure 4.
Methods to combine TL with ACS and AL.
Left: Combining TL and ACS; Right: Combining TL and AL.
Figure 5.
The TL+ACS algorithm, which uses TL to determine the optimal
and ACS to generate new training samples.
Figure 6.
Left: Color naming; Middle: Word reading; Right: Interference.
Table 1.
Mean and standard deviation of two performance measures in different scenarios.
Table 2.
The 29 features used by the kNN and SVM classifiers.
Figure 7.
Coefficients of the first two principle components of the 29 features.
Figure 8.
Performances of the four kNN classifiers on the 18 subjects for
. The horizontal axis shows
, and the vertical axis shows the testing accuracy on the
examples from the same subject.
Figure 9.
Mean and standard deviation (std) of the four kNN classifiers on the 18 subjects.
is the number of primary training samples generated in each iteration.
Table 3.
Average classification accuracy for the four methods.
Table 4.
Paired -test results (
) on classification accuracy.
Figure 10.
Mean and standard deviation of the percentage of primary training samples saved by TL, ACS, and TL+ACS over the baseline approach, when the kNN classifier is used.
is the number of primary training samples generated in each iteration.
Table 5.
Average percentage of saved primary training samples over the baseline method.
Table 6.
Paired -test results (
) on percentage of saved primary training samples.
Figure 11.
Performances of the four SVM classifiers on the 18 subjects for
. The horizontal axis shows
, and the vertical axis shows the testing accuracy on the
examples from the same subject.
Figure 12.
Mean and standard deviation (std) of the four SVM classifiers on the 18 subjects.
is the number of primary training samples generated in each iteration.
Figure 13.
Mean and standard deviation of the percentage of primary training samples saved by TL, ACS, and TL+ACS over the baseline approach, when the SVM classifier is used.
is the number of primary training samples generated in each iteration.
Figure 14.
Comparison of the kNN and SVM classifiers for
.