Fig. 2. Stimuli for current study.

Ten representative natural images from each of the categories we use in our experiments, arranged by the entropy of their associated human categorizations. Images were taken from the CIFAR-10 test set⁴⁹ and clustered into ten entropy bins, with the same minimum and maximum entropy used for all categories. Representative images were then chosen from each bin, for each category. Lower entropy means a higher degree of human categorization consensus.