An important hurdle to overcome before machine learning models can be reliably deployed in practice is identifying when samples are different from those seen during training, as the output for unexpected samples are often confidently incorrect, while not being identifiable as such. This problem is known as out-of-distribution (OOD) detection. A popular approach for the unsupervised OOD case is to reject samples with a high Mahalanobis distance with regards to the mean features of the training

more »

... ta. Recent work showed that the Mahalanobis distance can be thought of as finding the training data invariants, and rejecting OOD samples that violate them. A key limitation to this approach is that it is limited to linear relations only. Here, we present a novel method capable of identifying non-linear invariants in the data. These are learned using a reversible neural network, consisting of alternating rotation and coupling layers. Results on a varied number of tasks show it to be the best method overall, and achieving state-of-the-art results on some of the experiments.