InputXGradientExplainer<T>

Models & Types · AiDotNet.Interpretability.Explainers

Input × Gradient attribution explainer - multiplies input values by their gradients.

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

For Beginners

Input × Gradient is one of the simplest gradient-based attribution methods. It multiplies each input feature by its gradient to get an attribution score.

Intuition:

  • Gradient tells you: “If I change this feature, how much does the output change?”
  • But gradient alone doesn’t consider the feature’s current value
  • Input × Gradient says: “The attribution is both HOW MUCH the feature matters AND what its value is”

Formula: attribution[i] = input[i] × gradient[i]

Why multiply by input? Consider a feature x with gradient g:

  • If x = 0 and g = 100: The feature COULD matter a lot, but currently contributes nothing
  • If x = 10 and g = 0.1: The feature has high value but low sensitivity
  • x × g captures both aspects

Comparison to other methods:

  • Simpler than Integrated Gradients (just one gradient computation)
  • Less theoretically grounded than SHAP (doesn’t satisfy Shapley axioms)
  • Good as a quick baseline or sanity check
  • Can have issues with saturation (gradients near zero even for important features)

When to use:

  • Quick initial analysis
  • As a baseline to compare against more sophisticated methods
  • When computational resources are limited
  • For debugging (if Input×Gradient and SHAP disagree dramatically, investigate)

Constructors

Constructor Summary
InputXGradientExplainer(Func<Vector<T>, Vector<T>> predictFunction, Func<Vector<T>, int, Vector<T>> gradientFunction, int numFeatures, string[] featureNames, bool absoluteValue) Initializes an Input × Gradient explainer with custom functions.
InputXGradientExplainer(INeuralNetwork<T> network, int numFeatures, string[] featureNames, bool absoluteValue) Initializes an Input × Gradient explainer with a neural network.

Properties

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

Methods

Method Summary
InputXGradientExplanation<T> Explain(Vector<T> instance) Computes Input × Gradient attribution for an input.
InputXGradientExplanation<T> Explain(Vector<T> instance, int? targetClass) Computes Input × Gradient attribution for an input with a specific target class.
InputXGradientExplanation<T>[] ExplainBatch(Matrix<T> instances)
Vector<T> ComputeGlobalAttributions(Matrix<T> data)
void SetGPUHelper(GPUExplainerHelper<T> helper)