Left, Spearman correlation between the average human ADM and the ADMs of the models (n = 5). Right, All-TNNs ( alpha = 10 and alpha = 100) exhibit significantly better alignment with human ADMs than control networks. The error bars show 95% CIs. The group-level human noise ceiling was calculated using split-halves with Spearman–Brown correction. All comparisons are significant except between the LCN and All-TNNs ( alpha = 1) after BH-FDR correction (two-sided permutation test, all significant, P < 1.6 x 10 -2 ; see all P values in Supplementary Fig. 22a ). In the box plots in e and f , the midline represents the median, the bounds indicate the 25th and 75th percentiles, and the whiskers extend to the 5th and 95th percentiles.
