On the Importance of Feature Separability in Predicting Out-Of-Distribution Error

Open in new window