Fig. 3. Performance of IT population on a battery of orthographic processing tasks.

a Performance of linear decoders trained on the IT and V4 representations on invariant orthographic tests, grouped into letter identification (n = 20 tests), bigram identification (n = 8 tests) and invariant word classification. Performance of artificial representations sampled from layers of deep convolutional neural network model CORnet-S are shown in grey. Error bars correspond to SD across feature samples (neural sites or CORnet-S features). b Selectivity of individual IT sites over 29 invariant orthographic processing tests. The heatmap shows selectivity significantly different from zero overall pairs of neural sites and tests. The histogram above shows the number of behavioral tests (N_t) that each neural site exhibited selectivity for; neural sites are ordered by increasing N_t. The histogram on the right shows the proportion of neural sites exhibiting selectivity for each test; the behavioral tests are ordered alphabetically within each task group (letter identification in orange, bigram identification in cyan, and word classification in gray). Dashed line corresponds to proportion of tests expected from chance (alpha = 5%). c Comparison of V4 and IT linear decoder performance across all 29 invariant orthographic tests from a and b, separated by letter size (small, medium, large). Inset shows the average difference in performance between IT and V4 (mean ± SE, n = 29 tests; positive values corresponding to IT > V4), for each letter size. The shaded region corresponds to the “normal” range of letter sizes, over which humans exhibit substantial tolerance. IT linear decodes consistently outperform V4 across this range of stimulus sizes, suggesting that the observed differences between IT and V4 are not simply a consequence of stimulus size.