A Fresh Look at Sanity Checks for Saliency Maps

Existing explanation tests have reliability problems: The widely-used Model Parameter Randomisation Test (MPRT), which checks whether AI explanations actually respond to changes in a model's parameters, has methodological flaws that can lead to unreliable results—meaning we may be trusting explanations that aren't truly trustworthy.

Noise undermines current evaluation methods: The original MPRT is sensitive to random noise in measurements, which can distort whether an explanation method passes or fails the sanity check. The proposed "Smooth MPRT" addresses this by using sampling techniques to filter out noise and provide more stable, dependable results.

Better testing leads to better AI explanations: By improving how we evaluate saliency maps (visual explanations showing which parts of an image a model focuses on), organizations can more confidently deploy explanation methods in real-world applications where transparency matters—such as healthcare, finance, or criminal justice.

Simpler, more direct evaluation is possible: The "Efficient MPRT" modification simplifies the testing process by measuring explanation complexity directly rather than relying on potentially biased similarity measurements, making it easier for practitioners to validate their explanation methods without specialized expertise.

Trustworthiness requires rigorous testing standards: These refinements demonstrate that responsible AI deployment depends on having robust sanity checks for explanations—not just having explanations, but having confidence that those explanations are actually reliable and meaningful.