Layer-Wise Relevance Propagation

Layer-Wise Relevance Propagation starts with the prediction score and redistributes that relevance backward through the network until it reaches the input.

Unlike ordinary gradients, LRP uses propagation rules chosen for each layer type.

Layer-wise relevance propagation diagram — Source: Arras et al. (2019)

ℹ

Relevance conservation

Many LRP rules aim to conserve relevance between layers:

$\sum_i R_i^{(l)}\approx\sum_j R_j^{(l+1)}$

The output score is decomposed into signed or unsigned relevance assigned to lower-level units.

LRP gives precise control and can produce readable token or pixel attributions. That control comes with responsibility: propagation rules and stabilizing terms can change the result.

Document the rule used for each layer and verify that relevance is conserved as expected.

Checkpoint

Why is model architecture knowledge especially important for LRP?

←PreviousIntegrated GradientsAttribution-Based Explanations Next→Occlusion MethodsAttribution-Based Explanations