A General Framework for Unsupervised Anomaly Detection

Abstract

In this paper we present GenAD, a simple and generic framework for detecting examples that lie out-of-distribution for a given training set. The approach is based on first learning a semantic neighbourhood for each example in the training set and then training a binary discriminator to classify whether two examples come from the same or different semantic neighbourhoods. The framework is applicable to a wide range of anomaly detection problems, including visual, audio, and textual data. We are able to outperform previous approaches for anomaly, novelty, or out-of-distribution detection in the visual domain by a large margin. In particular we obtain AUROC >0.999 for the challenging task of detecting examples from SVHN as out-of-distribution given CIFAR-10 as in-distribution, without making use of label information.

Tasks

Reproductions