Learning Outlier-Aware Representation with Synthetic Boundary Samples

Out-of-distribution (OOD) detection provides an essential scheme to build a reliable deep learning system. Some of the previous works additionally acquired OOD samples for model training or hyperparameter tuning. However, OOD samples are usually missing in real-world scenarios and the prior knowledge of these samples is commonly unavailable. To cope with these issues, this paper presents a novel approach to synthesizing the informative samples near the boundary between in-distribution and OOD which sufficiently reflect OOD samples in training stage. These synthetic boundary samples are located right outside in-distribution (ID) where the latent variable is distributed by a multivariate Gaussian. Accordingly, this paper presents the outlier-aware representation learning which utilizes the synthetic boundary samples to train an OOD detector by only using the unlabeled ID data. The model is then trained to learn a compact decision boundary between ID and OOD samples. The experiments demonstrate that the proposed methods outperform state-of-the-art performance in presence of different OOD data.