Screen 1 of 3 - Read before you experiment

An explanation highlights evidence; it does not prove causality

Grad-CAM localizes influential regions in a convolutional feature map. Integrated Gradients attributes the prediction relative to a baseline. Occlusion measures how the score changes when image regions are hidden.

Different methods answer different questions and may disagree. First predict where the model should look, then compare the explanation with the object and its surrounding context.

Evidence levelAuthentic ResNet-18 predictions + precomputed explanation maps
Screen 2 of 3 - Experiment

Run, inspect, and compare

Follow the three guided moves above. Change one variable at a time so every visual change has a clear cause.

Continue to explanation reliability
Screen 3 of 3 - Consolidate

By the end of this lesson, you will be able to:

  • Interpret positive and negative feature attributions for an individual prediction.
  • Distinguish a local explanation from claims about global model behaviour.
  • Assess whether an explanation remains stable under small, label-preserving input changes.

Covers GradCAM, SHAP, and permutation importance. Discusses trust, compliance, and debugging. Notes limitations in resolution, cost, and accuracy.

 

GradCAM_demo