InfluenceFunctionExplainer<T>

Models & Types · AiDotNet.Interpretability.Explainers

Influence Function explainer for training data attribution.

Implements: IGPUAcceleratedExplainer<T>

For Beginners

Influence Functions answer the question: “Which training examples were most responsible for this prediction?”

This is different from feature attribution (which features matter) - instead it tells you which TRAINING DATA points matter. This is incredibly useful for:

Use Cases:

  • Debugging: Finding mislabeled training data
  • Data cleaning: Identifying harmful training examples
  • Understanding: Seeing which examples the model learned from
  • Fairness: Finding training data that causes biased predictions

How it works: Influence Functions use calculus to efficiently approximate: “What would happen to this test prediction if we removed a specific training example and retrained?”

Instead of actually retraining (which is expensive), we use the Hessian (second derivatives of the loss) to estimate the effect mathematically.

The math (simplified): influence(training_point) = (gradient_test) * (inverse_Hessian) * (gradient_train)

  • gradient_test: How the test loss changes with parameters
  • inverse_Hessian: How parameter changes propagate through the model
  • gradient_train: How this training point affected the parameters

Interpretation:

  • Positive influence: Removing this training point would HURT test performance
  • Negative influence: Removing this training point would HELP test performance
  • Large magnitude: This training point had a big effect

Constructors

Constructor Summary
InfluenceFunctionExplainer(Func<Vector<T>, Vector<T>> predictFunction, Func<Vector<T>, Vector<T>, T> lossFunction, Func<Vector<T>, Vector<T>, Vector<T>> gradientFunction, Matrix<T> trainingData, Vector<T> trainingLabels, InverseHessianMethod method, double damping, int maxIterations, int recursionDepth, double scale, int? randomState, Func<Vector<T>, Vector<T>, Vector<T>, Vector<T>> hvpFunction) Initializes a new Influence Function explainer with custom gradient function.
InfluenceFunctionExplainer(INeuralNetwork<T> network, Func<Vector<T>, Vector<T>, T> lossFunction, Matrix<T> trainingData, Vector<T> trainingLabels, InverseHessianMethod method, double damping, int maxIterations, int recursionDepth, double scale, int? randomState, Func<Vector<T>, Vector<T>, Vector<T>, Vector<T>> hvpFunction) Initializes a new Influence Function explainer.

Properties

Property Summary
bool IsGPUAccelerated { get; }
bool SupportsGlobalExplanations { get; } Gets whether this explainer supports global explanations.
bool SupportsLocalExplanations { get; } Gets whether this explainer supports local explanations.
string MethodName { get; } Gets the method name.

Methods

Method Summary
InfluenceFunctionResult<T> ComputeInfluence(Vector<T> testInput, T testLabel) Computes the influence of all training samples on a test sample.
SelfInfluenceResult<T> ComputeSelfInfluence() Computes the self-influence of each training sample.
TracInResult<T> ComputeTracIn(Vector<T> testInput, T testLabel, List<Matrix<T>> checkpointGradients) Computes TracIn-style influence using gradient checkpoints.
void SetGPUHelper(GPUExplainerHelper<T> helper)