Fig. 6

Plots for the first 10 datasets listed in Table 1. For each dataset, on the left, we display the robustness values of the training and test set ordered by size from least robust (left) to most robust (right) as blue and red curves, respectively. Since the test sets contain four times fewer elements than the training sets, we stretched the indices at which we plot the robustness values by a factor of four. The mean average curves over the 10 splits are indicated by dots and pluses, respectively, while the curves of the individual splits are shown in the same but more transparent colours. The right plots show histograms of the robustness values of the training sets. In both cases, we marked the thresholds corresponding to the 5%/10%/20% least robust training elements