IntegratedGradientsExplainer<T>

Models & Types · AiDotNet.Interpretability.Explainers

Integrated Gradients explainer for neural networks with gradient access.

Implements: IGPUAcceleratedExplainer<T>, IGlobalAttributionExplainer<T>, ILocalExplainer<T, IntegratedGradientsExplanation<T>>

For Beginners

Integrated Gradients is a method for explaining neural network predictions that satisfies two important mathematical properties:

  1. Completeness (Axiom of Completeness): The attributions sum up to the difference

between the prediction at the input and the prediction at a baseline (usually zeros).

  1. Sensitivity: If a feature differs between input and baseline and affects the output,

it gets a non-zero attribution.

How it works:

  • Start with a “baseline” (typically all zeros or a neutral input)
  • Create a path from baseline to your actual input
  • Integrate the gradients along this path
  • The result shows how much each feature contributed to moving from baseline to final prediction

Why use Integrated Gradients?

  • Theoretically sound (satisfies axioms that other methods don’t)
  • Works with any differentiable model
  • Attributions have clear meaning: contribution to prediction difference from baseline

Example: For an image classifier predicting “cat”:

  • Baseline: black image (all zeros)
  • Input: image of a cat
  • Integrated Gradients shows which pixels contributed most to the “cat” prediction

Constructors

Constructor Summary
IntegratedGradientsExplainer(Func<Vector<T>, Vector<T>> predictFunction, Func<Vector<T>, int, Vector<T>> gradientFunction, int numFeatures, int numSteps, Vector<T> baseline, string[] featureNames) Initializes a new Integrated Gradients explainer.
IntegratedGradientsExplainer(INeuralNetwork<T> neuralNetwork, int numFeatures, int numSteps, Vector<T> baseline, string[] featureNames) Initializes a new Integrated Gradients explainer from a neural network model.

Properties

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

Methods

Method Summary
IntegratedGradientsExplanation<T> Explain(Vector<T> instance) Computes Integrated Gradients attributions for an input.
IntegratedGradientsExplanation<T> Explain(Vector<T> instance, int outputIndex) Computes Integrated Gradients attributions for a specific output class.
IntegratedGradientsExplanation<T>[] ExplainBatch(Matrix<T> instances)
Vector<T> ComputeGlobalAttributions(Matrix<T> data)
void SetGPUHelper(GPUExplainerHelper<T> helper)