$\mathcal{G}$-Distillation: Reducing Overconfident Errors on Novel Samples

Counter to the intuition that unfamiliarity should lead to lack of confidence, current algorithms often make highly confident yet wrong predictions when faced with test samples from an unknown distribution different from training. Unlike all domain adaptation methods, we cannot gather an unexpected dataset prior to test. We propose a simple solution that reduces overconfident errors of samples from an unknown distribution without increasing evaluation time: train an ensemble of classifiers and then distill into a single model using both labeled and unlabeled examples. Experimentally, we investigate the overconfidence problem and evaluate our solution by creating and novel test splits, where are identically distributed with training and novel are not. We show that our solution yields more appropriate prediction confidences, on familiar and data, compared to single models and ensembles distilled on training data only. For example, we reduce confident errors in gender recognition by 94% on demographic groups different from the training data.