Fig. 4: Deliberation times reflected algorithm requirements.

a, Detail of the task schematic highlights the ‘deliberation time’, between the instance presentation and the first selection, and the ‘selection interval time,’ between the subsequent selections. b, Box plots showing the within-instance differences in average deliberation times between solutions classified as low or high complexity. Orange and brown box plots represent data from monkeys G and B, respectively. n = 201 and n = 270 instances for monkeys G and B, respectively. c, Box plots showing the within-instance differences in average selection intervals times between solutions classified as low or high complexity. Orange and brown box plots indicate data from monkeys G and B, respectively. n = 186 and n = 164 instances for monkeys G and B, respectively. d, Box plots showing the deliberation times for high-complexity solutions, divided according to whether the estimated thresholding parameter, t, successfully reduced (deep purple) or failed to reduce (magenta) the size of the combinatorial search space (Methods). n = 2507 and n = 5165 trials for monkeys G and B, respectively. e, Image of a training strategy for creating RNNs that simulated low-complexity (green) or high-complexity (purple) algorithms. The blue circles represent the 20 deliberation steps, and the target solutions are defined by the low- or high-complexity algorithms. f, Scatter plot for normalized performance when the networks were prematurely stopped during the deliberation steps. Normalized performance was the performance of the RNNs at each time step divided by the performance of the RNNs that underwent the entire deliberation period. g, Histograms show the frequency of considering different numbers of unique combinations during the deliberation steps. Green and purple histograms represent data from RNNs trained to mimic low- and high-complexity algorithms, respectively. n = 462 instances. Box plots show the median (line), quartiles (boxes), range (whiskers) and outliers (+).