Extended Data Fig. 10: Dopamine transients represent prediction errors better than mismatch errors.
From: Natural behaviour is learned through dopamine-mediated reinforcement
a, Top, comparison of the PPE and ME models with different numbers of performance history terms (see Methods). Each trace is a single syllable model comparison (blue dot, the minimum, negative Akaike information criterion, ∆AIC, the PPE model’s best fit relative to the ME model). The magnitude and sign of ∆AIC indicate the relative superiority of one model over another. 25/25 syllables were included in the ME/PPE comparison because all syllables had at least one ME or PPE linear regression onto dopamine with a significant R2 value (computed using the fitlm function in MATLAB). 22/25 syllable traces have a negative minimum (blue dot excluded from the three syllable traces which are always above 0; that is, the ME model always outperforms the PPE model). Blue line indicates the average best number of history terms across syllables (including the three syllables in which no. history terms = 0). Bottom, summary plot of number of history terms in best selected model from the top traces. 3/25 models had no PPE models which improved fit to dopamine over the ME model (shown with open circle). b, Top, as in a, but applying ∆MSE as a secondary model comparison metric (see Methods). As in a, the sign and magnitude of ∆MSE indicates the relative superiority of the ME versus PPE model. Negative values indicate that the PPE model outperforms the ME model. The ∆AIC and ∆MSE metrics found the average best number of history terms across syllables to be n = 11 and n = 12, respectively. Bottom, plotted similarly to a for ∆MSE.
