Information Technology Reference
In-Depth Information
Block Size
Correctly
Classified Instances
Incorrectly
Classified Instances
10
70.3%
29.7%
3
60.6%
39.4%
Fig. 3.5
The results show that RSVP can be used successfully to build a separable data set for
training a P300 classi
er. The single trial results (Block Size 1) show 70 % accuracy
were removed using a cascading FIR
filter, which was chosen to avoid phase
distortion. From successful trials, the windows from 220 ms to 440 ms after the
oddball presentation were collected (set A). Trials where the participant was unsure
or gave an incorrect answer were collected in set B. Outliers were rejected from
these sets, by removing windows that were more than three standard deviations in
Euclidean distance from the mean, leaving 180 trails in set A and 40 trials in set B.
Data sets were created for classi
cation, by creating each training example as an
average over a number of windows. Drawing on set A, training sets were creating
of 180 examples, using average sizes of 10, 3 and 1 (single trial). The corre-
sponding sets of negative examples were creating using averages of windows from
random time points in set B. These sets were used to train a bagging classi
er, using
a random forest as a sub-classi
er. The results from tenfold validation tests were as
follows (Fig.
3.5
).
These results demonstrate that using machine learning, it is possible to create a
P300 classi
er using consumer hardware. This has exciting implications for the
future usability of BCI systems. Given the interaction designs detailed in this
chapter, it should be possible in the near future to create highly accessible, low-cost
BCMI systems, and it is this goal that we continue to pursue.
3.9
Questions
1. What is an ERP?
2. How are ERPs different to spontaneous brainwave potentials?
3. How are ERPs different to SSVEPs?
4. What are the main drawbacks of the ERP technique for musical interaction?
5. What is the purpose of the P300 averaging technique?
6. Describe one way that ERP techniques could be used to get information about
how listeners experience music.
