Fig. 5: A dual-controller model predicts the effect of experimental manipulations.
From: Dopaminergic action prediction errors serve as a value-free teaching signal
a, Schematic of the network model. b, Top, task performance across learning for the full dual-controller model (combined), the value-based controller, or the value-free controller. Bottom, differences in performance between the combined model and the value-based model (12 random agents selected for each). c, Change of the model weights for a reward association (high tone→left action), as means for 100 agents. Vertical lines indicate inactivation time points in d. d, Performance levels before and after (as the mean after ten trials) model inactivations of the TS or the actor networks. e, Schematic of the experimental approach for acute inhibition of D1 SPNs or D2 SPNs in the TS. f, Quantification of the contralateral bias on opto-stimulated trials (Methods) for each session as a function of session’s performance. Error bars represent 95% confidence intervals. Lines show the mean and s.d. of linear fits for each mouse in each dataset. D1-Arch P = 0.004; A2A-Arch P = 7.6 × 10−4; Methods; n = 8 mice. g, Proportion of significantly biased sessions in f as a function of performance. h, Final weights in the TS for each sound–action association.
