Detecting Out-of-distribution Samples via Variational Auto-encoder with Reliable Uncertainty Estimation

In unsupervised learning, variational auto-encoders (VAEs) are an influential class of deep generative models with rich representational power of neural networks and Bayesian methods. However, VAEs suffer from assigning higher likelihood to out-of-distribution (OOD) inputs than in-distribution (ID) inputs. Recent studies advise that the deep generative models with reliable uncertainty estimation is critical to a deep understanding of OOD inputs. Meanwhile, noise contrastive prior (NCP) is an emerging promising method for obtaining uncertainty, with the advantages of easy to scale, being trainable, and compatibility with extensive models. Inspired by these ideas, We propose an improved noise contrastive prior (INCP) to acquire reliable uncertainty estimate for standard VAEs. By combining INCP with the encoder of VAE, patterns between OOD and ID inputs can be well captured and distinguished. Our method outperforms standard VAEs on the FashionMNIST and CIFAR10 datasets. We also demonstrate the preferred robustness of our model by the extensive experiments on anomaly detection tasks.