dda99de58ff020cfb57fec1404c97003-Supplemental.pdf

WhereR is number of possible classes,N is overall number of instances of all classes.fti is the approximated probability of the forecastoti in one hot encoding. Theobvious disadvantage toDeep Ensembles isthatitrequiresN times as long to train and run inference as its base model. We can see consistently that our model had stronger calibration across all models and metrics, including models known to be well calibrated like LeNet [22].