Fig. 3

Hold-out error rates and computational demands as functions of \(\textit{S}_{\textit{c}}\), \(\textit{P}\!_{\textit{v}}\) and \(\textit{R}\). These plots visualize the most important results displayed in Table 2. Plots a–c show the average classification errors obtained with the individual hold-out method with equal a priori probabilities \(\textit{p}_{\textit{s}}\!=\!\textit{p}_{\textit{d}}\) =50 % (14) for different values of the cutoff scale \(\textit{S}_{\textit{c}}\) and the percentage of variance \(\textit{P}\!_{\textit{v}}\) explained by PCA. The corresponding error rates obtained by using the knowledge of the real oddball probabilities \(p_s\) = 86.4 % and \(\textit{p}_{\textit{d}}\) = 13.6 % (15) are displayed in the plots d–f. The approximate processing time as a function of \(\textit{S}_{\textit{c}}\), or the respective cutoff frequency \(\textit{f}_{\textit{c}}\) (22), is displayed in plot g. Plot h shows how the number of the non-zero CWT matrix elements, measured in millions (mln), and the respective memory usage, measured in megabytes (MB), depend on \({\textit{S}}_{\textit{c}}\) or \(\textit{f}_{\textit{c}}\). Plot i shows how both processing time and memory usage increase as a function of the log-grid sampling rate R (40)