Abstract

This presentation discussed a new approach that was published in ICLR2019, that allows incorporating out-of-distribution (OOD) example detection capabilities to deep networks, but adding a single loss term to the training loss function. The proposed method may have a number of potential uses in critical fields where out-of-distribution samples arise often, and are misclassified with a very high confidence score by a classifier. This method trains the network to be ‘unsure’ when it sees OOD samples, and predict conservatively.