Fig. 4: Human performance in the binary decision task.
a, Accuracy does not differ between early and late training trials for any level of diagonality in Kacc. The lines represent the means, and the shaded intervals denote 90% confidence intervals. b, However, in Krew there is a significant interaction such that accuracy increases more for oblique than for cardinal orientations as predicted by a decrease in the q parameter to implement a fitness-maximizing code. c, Fitting the early and late training decisions separately in Krew and Kacc showed that decreases in q for Krew training were greater than for Kacc training. In each panel, the data from Kacc are shown in blue, while those from Krew are shown in red. The violin plots in c show the posterior distributions for the group-level estimates of q in late minus early trials for Kacc and Krew training sessions. The horizontal black lines indicate the medians of the posterior distributions, and the width of the violin plot represents their density.
