Explainable Deep Learning Models

Faculty: Lyle Ungar

*Importances of individual features are deceptive because of correlations; Use Shapley values to instead measure the importance of sets of features

Opportunity: Explain deep learning models that predict everything from politeness in different languages to outcomes in surgery.


Challenge: No one (or 10) words or medical device measurements drive the prediction, and different words and measurements are used in different countries and hospitals.