Figure 6.

Model Performance versus Supervision Type on Auxiliary (Text) and Target (Image or Signal) Domains

We analyzed the performance of majority vote (MV) of the labeling functions (LFs), a generative model trained on these LFs (GM), an LSTM trained to map the raw text to the MV output (DM-MV), an LSTM trained to map the raw text to the GM output (DM-GM), and hand-labeled full supervision (FS). Text model performance is evaluated on the development set, as coverage and ROC-AUC on this set are used in the cross-modal data programming heuristic optimizer. Target modality performance is evaluated on the held-out test set. We present results for CXR (A), EXR (B), HCT (C), and EEG (D). Note that MV text results and coverage results are deterministic. Error bars are 95% confidence intervals from five runs with different random seeds.