Click each sample to see how badly an autoencoder trained on normal data rebuilds it. The unusual one scores a high error and is flagged.
Train only on normal data. Anything the model rebuilds badly is something it has never seen — no anomaly labels required.
