Generative Probabilistic Novelty Detection with Isometric Adversarial Autoencoders

Abstract

Learning the manifold of a complex distribution is a fundamental challenge for novelty or anomaly detection. We introduce a revised learning and inference procedure that takes into account a key underlying assumption made by the framework of generative probabilistic novelty detection. The traditional framework implies the ability to not only learn the manifold of the generative distribution of inliers but also to compute non-linear orthogonal projections onto this manifold from the ambient space. We augment the original training with priors that endow the model with this property, and prove that inference becomes easier and computationally more efficient. We show experimentally that the new framework leads to improved and more stable results.