Understanding Failures in Out-of-Distribution Detection with
                 Deep Generative Models

                  Lily H. Zhang 1 Mark Goldstein 1 Rajesh Ranganath 1

              Abstract                   distribution, estimated via a model (Bishop, 1994).
    Deep generative models (DGMs) seem a natu-            Deep generative models (DGMs) estimate complex distri-
    ral fit for detecting out-of-distribution (OOD) in-       butions from often high-dimensional inputs and produce
     ...