Reviews: Generative Probabilistic Novelty Detection with Adversarial Autoencoders

The problem of the study is outlier detection given only inlier training data. The high-level approach is to learn a density estimation function over the training data, and then filter out outliers using a learned threshold \gamma. They approximate the density function through a decomposition over the tangent space of the learned manifold near each given sample. To learn the manifold structure the authors use a variation of adversarial autoencoders. The evaluation is performed on MNIST, FashionMNIST, and COIL against a few baselines. Overall, the paper is very well-written and easy to follow -- the presentation progresses coherently.