OcclusionExplainer<T>

Models & Types · AiDotNet.Interpretability.Explainers

Occlusion explainer for image and sequential data interpretation.

Implements: IGPUAcceleratedExplainer<T>, ILocalExplainer<T, OcclusionExplanation<T>>

For Beginners

Occlusion is a simple but powerful explanation technique. The idea is: systematically hide different parts of the input and see how the prediction changes.

How it works:

  1. Take your input (e.g., an image)
  2. Place a “patch” over one part of the input (occlude it)
  3. See how the model’s prediction changes
  4. Move the patch and repeat
  5. The result is a map showing which regions matter most

Intuition: If covering a region causes the prediction to drop significantly, that region was important for the prediction.

Use cases:

  • Understanding which parts of an image a classifier looks at
  • Finding what regions of a medical scan led to a diagnosis
  • Debugging models that use spurious correlations (e.g., looking at background)

Advantages:

  • Very simple and intuitive
  • Model-agnostic (works with any model)
  • Easy to visualize

Disadvantages:

  • Can be slow (many forward passes required)
  • May not capture feature interactions well
  • Results depend on occlusion patch size

Constructors

Constructor Summary
OcclusionExplainer(Func<Tensor<T>, Tensor<T>> predictFunction, int[] windowShape, int[] strides, T baselineValue, OcclusionShape shape) Initializes a new Occlusion 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
List<OcclusionExplanation<T>> ExplainBatch(List<Tensor<T>> inputs, int? targetClass) Explains a batch of inputs.
OcclusionExplainer<T> ForImages(Func<Tensor<T>, Tensor<T>> predictFunction, int patchHeight, int patchWidth, int? strideHeight, int? strideWidth, T baselineValue, OcclusionShape shape) Initializes an Occlusion explainer for 2D images.
OcclusionExplanation<T> Explain(Tensor<T> input, int? targetClass) Explains a single input by computing occlusion sensitivity.
OcclusionExplanation<T> Explain(Vector<T> instance) Explains a single vector input.
OcclusionExplanation<T>[] ExplainBatch(Matrix<T> instances) Explains a batch of inputs.
void SetGPUHelper(GPUExplainerHelper<T> helper)