Autoencoders

Best for: Complex high-dimensional anomalies

How it works

Trains an encoder $E$ to compress each input into a low-dimensional code and a decoder $D$ to reconstruct it, $\hat{x}=D(E(x))$, so the model learns the manifold of normal data. At inference the reconstruction error $\mathcal{R}(x)=\|x-\hat{x}\|^2$ is large for inputs unlike anything seen in training, because the bottleneck cannot reproduce them faithfully. A threshold $\tau$ on $\mathcal{R}$, often a high quantile of validation errors, separates anomalous from normal observations.

Common fields

Images · network traffic · medical scans