Fig. 3

Topological changes during the middle period. A A schematic depiction of two topological extremes: on the left is a segregated network, with tight-knit communities that are weakly inter-connected—this network would be characterized by high Q, and would have more nodes with high module-degree z-score (MZ) than nodes with high participation coefficient (PC); on the right is an integrated network, which has stronger connections between nodes in different communities, and hence a lower Q, and more nodes with high PC than nodes with high MZ; B participation coefficient (PC) of input layer nodes at training epoch 30; C module-degree z-score (MZ) of input layer at training epoch 30; D digit information, I_D = MI(pixel_On, class); E Pearson’s correlation, r, between I_D and PC (red) and MZ (blue) across first 30 training epochs. Black lines represent the upper and lower bounds (95th and 5th percentiles) of a permuted null dataset (10,000 iterations) and colored bars represent learning periods; F I_H = MI(node, class) for HL1 (blue) and HL2 (orange) edges—note that both subnetworks increase I_H during the middle period, but that the late period dissociates the two subnetworks