Latent Distribution of Variational Autoencoder for Out-of-Distribution Detection

Abstract

Deep generative models, which can obtain the likelihood of each data point, are often used for out-of-distribution detection. However, deep generative models may assign a higher likelihood to out-of-distribution data. VAE is one of the deep generative models suffering from this problem. The typical prior distribution is the standard normal distribution in VAEs. We hypothesize that the latent variables of out-of-distribution data are concentrated near the origin. To solve this problem, this paper proposes an alternative prior distribution of the latent variables in VAEs. The proposed prior distribution has a low probability density near the origin. Experiments on Fashion-MNIST and MNIST show that the proposed method improves the performance of out-of-distribution detection compared to existing methods.