Skip to main content
Fig. 5 | Journal of Cheminformatics

Fig. 5

From: Evaluating uncertainty-based active learning for accelerating the generalization of molecular property prediction

Fig. 5

(Upper Left) Test set RMSE values increase for molecules in each PC value bin when training molecules from that bin are removed from training (blue solid line) relative to when the full training set is used (blue dashed line). RMSE values for molecule types which are not removed from the training set are not affected (green lines). (Upper Right) Uncertainty estimates generated using a density approach show that changes in uncertainty follow a similar pattern to changes in error. (Lower) The correlation of \(\Delta\)Error and \(\Delta\)UQ across all PCA components and bins is 0.818

Back to article page