LayerwiseRelevancePropagationExplainer<T>

Models & Types · AiDotNet.Interpretability.Explainers

Layer-wise Relevance Propagation (LRP) explainer for neural networks.

Implements: ILocalExplainer<T, LRPExplanation<T>>

For Beginners

LRP explains neural network predictions by propagating “relevance” scores backward from the output to the input.

Key idea:

  • Start with the prediction score as the total “relevance”
  • At each layer, distribute relevance to neurons in the previous layer
  • Continue until you reach the input features
  • Each input feature gets a relevance score

Conservation principle:

  • The total relevance is conserved at each layer
  • Relevance in = Relevance out (like energy conservation)
  • This means attributions sum to the prediction value

LRP rules (how to distribute relevance):

  • LRP-0 (Basic): Distribute proportional to contribution
  • LRP-ε (Epsilon): Adds stability with small epsilon
  • LRP-γ (Gamma): Emphasizes positive contributions
  • LRP-αβ: Separately handles positive and negative contributions

When to use LRP:

  • You want to understand which inputs were “responsible” for the output
  • You need attributions that sum exactly to the prediction
  • You want a principled way to handle different layer types

Example: For image classification:

  • LRP shows which pixels contributed positively or negatively
  • Red = positive relevance (supported the prediction)
  • Blue = negative relevance (contradicted the prediction)

Constructors

Constructor Summary
LayerwiseRelevancePropagationExplainer(Func<Vector<T>, Vector<T>> predictFunction, Func<Vector<T>, Vector<T>[]> getLayerActivations, Func<int, Matrix<T>> getLayerWeights, int numFeatures, int[] layerSizes, string[] featureNames, LRPRule rule, double epsilon, double gamma) Initializes a new LRP explainer.

Properties

Property Summary
bool SupportsGlobalExplanations { get; }
bool SupportsLocalExplanations { get; }
string MethodName { get; }

Methods

Method Summary
LRPExplanation<T> Explain(Vector<T> instance) Computes LRP relevance scores for an input.
LRPExplanation<T> Explain(Vector<T> instance, int outputIndex) Computes LRP relevance scores for a specific output.
LRPExplanation<T>[] ExplainBatch(Matrix<T> instances)