#clever-hans

Question 1: Single choice

A classifier separates photos of huskies from photos of wolves with high accuracy, but LIME explanations of its "wolf" predictions mostly highlight the snowy background instead of the animal. What is the most likely cause?

Training photos of wolves showed snow, so snow became a shortcut1
LIME cannot handle images, so its highlights carry no information2
The model is underfitting and would focus on the animal after more epochs3
The explanation is unfaithful, because the model really looks at the animal4

Timed test