Figure 4.

Mean Neural Network ROC-AUC Score versus Dataset Size Using Full Supervision and Cross-Modal Data Programming

Results are presented for chest radiographs (CXR) (A), extremity radiographs (EXR) (B), head CT (HCT) (C), and electroencephalography (EEG) (D). DP, cross-modal data programming; FS, full hand-labeled supervision. In each case, the performance of models trained using cross-modal data programming improve as additional unlabeled data are added, and in several cases exhibit scaling properties very similar to those of the fully supervised model as additional labeled data are added. Error bars (dashed lines for FS, shaded region for DP) represent 95% confidence intervals from five training runs with different random seeds.