*3.3. Classifier Training Time*

In order to evaluate the training time of the investigated classifiers, the cross-validation scheme was run four times for each subject, each time with an increasing number of EEG channels retained in the analysis, to explore the scalability of each classifier for analyses with higher spatial resolutions. The temporal resolutions were not varied, but we expect that increasing the temporal resolution has a similar effect on training time, since the training times for the STBF-based classifiers are primarily dependent on the number of parameters in their respective covariance matrix estimators, as evidenced by the complexity calculations in Section 4.2. Figure 4 shows the measured training times. These results were obtained using a laptop with an Intel® Core™ i7-8750H CPU (Intel Corporation, Santa Clara, CA, USA) and 16 GB of RAM.

**Figure 4.** Median fold training time for different classifiers at different spatial resolution levels evaluated over all training folds for all subjects. Error bars represent standard deviation. The training time of STBF-SHRUNK increased more steeply with resolution compared to STBF-STRUCT. All STBF classifiers were able to be trained significantly faster than xDAWN+RG.

Figure 4 shows that the training time of STBF-STRUCT increased less steeply than that of STBF-SHRUNK and STBF-EMP. The training time of all three STBF-based classifiers was much lower than that of xDAWN+RG, which appears nearly constant when using 4, 8, 16, or 32 channels.
