Outlier detection is an important field in data mining. For high-dimensional data the task is particularly challenging because of the so-called “curse of dimensionality”: The notion of neighborhood becomes meaningless, and points typically show their outlying behavior only in subspaces. As a result, traditional approaches are ineffective. Because of the lack of a ground truth in real-world data and of a priori knowledge about the characteristics of potential outliers, outlier detection should be considered an unsupervised learning problem. In this paper, we examine the usefulness of unsupervised artificial neural networks – autoencoders,self-organising maps and restricted Boltzmann machines – to detect outliers in high-dimensional data in a fully unsupervised way. Each of those approaches targets at learning an approximate representation of the data. We show that one can measure the “outlierness” of objects effectively, by measuring their deviation from the learned representation. Our experiments show that neural-based approaches outperform the current state of the art in terms of both runtime and accuracy.