Skip to content

Deep SVDD — One-Class Anomaly Detection

Deep Support Vector Data Description (Deep SVDD) learns a compact representation of "normal" data, then flags anything that lands far from the centre at inference time. It's the gradient-trained sibling of RX anomaly detection — slower to train, but able to capture non-Gaussian backgrounds.

This tutorial walks through training Deep SVDD end-to-end on a hyperspectral dataset and rendering anomaly heatmaps for each frame.

Run the example:

What you'll learn

  • How the gradient training stage consumes a statistically-initialised pipeline.
  • Configuring the representation dimension, soft boundary, and unfreeze schedule.
  • Saving the trained pipeline and trainrun config so the run is reproducible via restore-trainrun.

When to reach for Deep SVDD

  • The background is non-Gaussian or multi-modal (RX would struggle).
  • You have enough "normal" data to train on and no labelled anomalies.
  • You're willing to spend training time to get a richer anomaly signal than statistical methods deliver.