Fig. 7. RNNs’ naïve weights at the end of the training step and their subsequent rates of change due to reward feedback during the simulation of our task.

a–c Plotted is the average strength of the naïve weights from feature-encoding, conjunction-encoding, and object-identity encoding populations to eight types of recurrent populations (indicated by the inset diagrams explained in Fig. 4b) (a), naïve weights between eight types of recurrent populations (b), and naïve weights from four types of excitatory recurrent populations to the output population (c). d Plotted is the average rate of value-dependent changes in the connection weights from feature-encoding, conjunction-encoding, and object-identity encoding populations to recurrent populations with plastic sensory input, during the simulation of our task. Asterisks and plus sign indicate two-sided and one-sided significant rates of change (sign-rank test; P < 0.05), respectively. Hatched squares indicate connections with rates of change that were not significantly different from zero (two-sided sign-rank test; P > 0.05). Highlighted rectangles in cyan, magenta, and red indicate the values for input from sensory units encoding the informative feature, the informative conjunction, and object-identity, respectively. e Plot shows the average rates of change in connection weights between recurrent populations. Conventions are the same as in d. Source data are provided as a Source Data file.