9f6992966d4c363ea0162a056cb45fe5-AuthorFeedback.pdf

Find way of unifying6 benchmarks. R1: Only CNNs trained on ImageNet were used/ R2: Include models trained on Stylized-ImageNet. Model shape bias predicts human-CNN error consistency for cue conflict stimuli,20 indicating that networks basing their decisions on object shape (rather than texture) make more human-like errors:21 modelshapebias(%) 20.5 21.4 34.7 81.4 human-CNNconsistency(κ).066.068.098.195 R4: Aggregating the classification probability by arithmetic mean may not be optimal.We have now included46 a principled derivation showing that the arithmetic mean is, perhaps counterintuitively, optimal.