Random Logit Metric¶

Metrics: Randomization

Random Logit Invariance metric tests whether explanations change when the target logit is randomized to a different class. This is a sanity check to verify that explainers are sensitive to the target label.

Quote

We propose sanity checks for saliency methods. [...] We find that some widely deployed saliency methods are independent of both the data the model was trained on, and the model parameters.

-- Sanity Checks for Saliency Maps (2018)¹

For each sample \((x, y)\):

Compute explanation for the true class \(y\)
Randomly draw an off-class \(y' \neq y\)
Compute explanation for \(y'\)
Measure SSIM (Structural Similarity Index) between both explanations

Info

A low SSIM indicates that explanations are sensitive to the target label (desirable if we expect class-specific explanations).

Score Interpretation¶

Lower scores are better: A low SSIM means the explanations change significantly when the target class changes, indicating the explainer is properly sensitive to the target.
Values range from -1 to 1, where 1 means identical explanations.
High SSIM values suggest the explainer may not be faithfully explaining class-specific features.

Example¶

from xplique.metrics import RandomLogitMetric
from xplique.attributions import Saliency

# load images, labels and model
# ...
explainer = Saliency(model)

metric = RandomLogitMetric(model, inputs, labels)
score = metric.evaluate(explainer)

Warning

This metric requires one-hot encoded labels with shape (N, C) where C is the number of classes.