Review for NeurIPS paper: Towards Maximizing the Representation Gap between In-Domain & Out-of-Distribution Examples

Weaknesses: Given that the major contribution of the paper is to improve detection of misclassified and OOD samples, it is a bit disappointing that the misclassification results in Table 2 and 8 are comparable at best with competing methods. In Fig 5, I'm curious what is the transform applied to the y-axis? L518 of the Appendix says: "In practice, since we do not know the characteristics of the OOD test examples, it may not be suitable to use a binary classifier for OOD detection tasks." This valid criticism also applies to the proposed method, since examples of outliers are used. Using binary classifiers has been shown to be very effective for highly practical tasks, for example, [3].