CNN Interpretability with LIME

PrototypeResearch

Model-agnostic feature-attribution pipeline across training checkpoints. With Marlon Dammann.

Superpixel attribution: which regions moved the predictionSample data

p = 0.94 cat9400%

p = 0.81 dog8100%

p = 0.63 bird6300%

p = 0.41 deer4100%

p = 0.88 ship8800%

p = 0.29 frog2900%

Warm overlay = regions that raised the predicted class. Tiles are schematic, not sample images.

PrototypeAn interface sketch of this system, not a screenshot. Every figure, name and record in it is invented.