Anomaly Detection and a Pointer to Variational Autoencoders

Every generative model built in this topic implicitly answers a density question: how plausible is this point under the fitted model? Mixture of Gaussians and EM fits a density directly, which makes it a natural tool for anomaly detection — flagging points that the model finds implausible. This note closes out the unsupervised-learning topic with that application, its precision/recall trade-off, model-selection via information criteria, and a forward pointer to variational autoencoders — a genuinely more expressive generative model that this foundational unit only surveys at the pointer level.