Fig. 4: Metacognition correlates with learning to use latent brain activity.
From: Unconscious reinforcement learning of hidden brain states supported by confidence
a Across-subject correlation between the baseline (minimal) gambling performance attained in sessions 1 and 2, and individual metacognitive ability (how well one’s confidence tracks discrimination accuracy). Metacognitive ability was computed with independent behavioural data from the decoder construction session (session 0, see ‘Methods'). Pearson correlation (n = 18), two-sided p value. b Proportion of optimal actions plotted by confidence level. The performance was measured as the proportion of trials in which the subject chose the action more likely to be rewarded, given the latent state. Significance was assessed with linear mixed effects models (two-sided p values, uncorrected). c Discrimination accuracy, as leftward vs. rightward motion discrimination, plotted by confidence level. The correctness of the response was based on the output of the decoder. Significance was assessed with linear mixed effects models (two-sided p values, uncorrected). For all plots, coloured dots represent individual subjects, grey bars the mean, error bars the SEM. The experiment was conducted once (n = 18 biologically independent samples). b **(interaction) P = 0.0017, **(session 2) P = 0.004, ***(session 3) P = 4.68 × 10−7; c ***(interaction) P = 4.68 × 10−6, †(session 2) P = 0.078, ***(session 3) P = 7.03 × 10−14.
