Unsupervised Out-of-Distribution Detection by Restoring Lossy Inputs with Variational Autoencoder

Deep generative models have been demonstrated as problematic in the unsupervised out-of-distribution (OOD) detection task, where they tend to assign higher likelihoods to OOD samples. Previous studies on this issue are usually not applicable to the Variational Autoencoder (VAE).