Improving the out-of-distribution performance of score-based generative models via self-supervision

Anonymous

Keywords: Out-of-distribution detection, Score-based generative models

TL;DR: We find existing score-based OOD detection methods are vulnerable to OODs that have similar textures and propose method to overcome this.

Abstract: In this work, we first examine the efficacy of score-based generative models (SGMs) for out-of-distribution (OOD) detection. We show previously proposed OOD detection metrics based on SGMs fail to address OODs that share similar textures but different object shapes. Based on the observation, we construct RotNCSN, a novel OOD detection method based-on the score matching and data augmentation. RotNCSN first applies random rotation to the perturbed data and forces its output to be rotation-invariant. Therefore, RotNCSN becomes more shape-aware. Experiment results show that RotNCSN consistently improves over the baseline metric based on the SGMs. Furthermore, RotNCSN also achieves competitive OOD detection performance in the FashionMNIST domain.

Student Paper: No