Accumulated local effects
Inspect accumulated local changes in predictions across a tabular feature.
Testing methods
72Accuracy
Measure how often predicted class labels match the reference labels.
Additive
Add a configured amount to each tabular feature value.
Annotation consistency
Compare class labels and annotation counts on exact duplicate images.
Annotation validity
Check that annotations can be read and refer to existing images and classes.
AP@0.50
Measure detection average precision at an IoU threshold of 0.50.
AP@0.75
Measure detection average precision at an IoU threshold of 0.75.
Aspect ratios
Find very wide or very tall images.
Blur
Blur images and compare results with the baseline.
Blur
Find images that lack sharp edges.
Brier gap
Compare Brier scores across groups.
Brightness
Change image brightness and compare results with the baseline.
Class balance
Compare how many images contain each declared class.
Contrast
Change image contrast and compare results with the baseline.
Contrast
Find images with a narrow range of tones.
Deletion and insertion AUC
Measure prediction changes as attributed image regions are removed or inserted.
Demographic parity ratio
Compare positive prediction rates across groups as a ratio.
Dice
Measure Dice overlap between predicted and reference masks.
Duplicates and split leakage
Find exact and near-duplicate images, including copies across dataset splits.
ECE gap
Compare expected calibration error across groups.
Energy pointing game
Measure how much attribution energy falls inside a reference region.
Equal opportunity difference
Compare true positive rates across groups.
Equalized odds difference
Compare true positive and false positive rates across groups.
Exposure
Find images that are mostly dark or washed out.
1–24 of 72 results