Fig. 2. The axis of object space for objects’ real-world size.

a Among all 50 PCs, only the removal of PC2 significantly reduced the correspondence between the RSM of Conv4’s responses and that of the ideal observer (Left). This is demonstrated by the RSM of Conv4’s residual responses after regressing out response variance aligned to PC2, which no longer showed sensitivity to objects’ size (Right). b The mapping of objects’ real-world size to PC2 was best described as a common logarithm function (left), among all functions tested (right). The color indicates size rank. c The removal of the response variance aligned to PC2 significantly impaired AlexNet’s performance on object recognition. S small objects, B big objects. Pow power, LN common logarithm.