Fig. 6.

Only a few labels are required to train an accurate model. Some labeled genomic windows were set aside as a test set, then models were learned for each training set size, for two different random orderings of the training set (mean line and min/max band). A circle on the right shows the test error of the model trained with the maximum number of windows. It is clear that in each dataset, the model-specific maximum accuracy is achieved after only 2–6 labeled windows (several dozen labels)