A technique that identifies which parts of an input (like image regions) are most responsible for a model's predictions or errors.