Bayesian Variational Autoencoders for Unsupervised Out-of-Distribution Detection

Despite their successes, deep neural networks may make unreliable predictions\nwhen faced with test data drawn from a distribution different to that of the\ntraining data, constituting a major problem for AI safety. While this has\nrecently motivated the development of methods to detect such\nout-of-distribution (OoD) inputs, a robust solution is still lacking. We\npropose a new probabilistic, unsupervised approach to this problem based on a\nBayesian variational autoencoder model, which estimates a full posterior\ndistribution over the decoder parameters using stochastic gradient Markov chain\nMonte Carlo, instead of fitting a point estimate. We describe how\ninformation-theoretic measures based on this posterior can then be used to\ndetect OoD inputs both in input space and in the model's latent space. We\nempirically demonstrate the effectiveness of our proposed approach.\n