Gradient Tutorials

Pipelines that include trainable parameters fit by backpropagation. They consume a statistically-initialised pipeline (Phase 1) and refine it through gradient descent (Phase 2). More expressive than purely statistical pipelines, at the cost of training compute.

In this section

• Deep SVDD — one-class anomaly detection that learns a compact representation of "normal" data. The gradient-trained sibling of RX.
• AdaCLIP — vision-language anomaly detection coupling a frozen CLIP backbone with a small trainable adapter. Three recipes: PCA-reduced baseline, Concrete channel selector, and DRCNN reducer.

See also

• Concepts → Training — the two-phase training model.
• Statistical tutorials — strong baselines, no gradient steps.