However, current approaches can only model distributions for which training samples are directly accessible, which is not the case in many real-world tasks.

There are a few approaches to building deep anomaly detection objectives or models.