Layer-wise relevance propagation (LRP)

Backward-propagates a model's output relevance score through each layer using conservation-preserving redistribution rules, attributing the prediction to specific input regions/features without requiring a second forward pass or an explicit direction/subspace.