The curious case of dopaminergic prediction errors and learning associative information beyond value.

Abstract

Transient changes in the firing of midbrain dopamine neurons have been closely tied to the unidimensional value-based prediction error contained in temporal difference reinforcement learning models. However, whereas an abundance of work has now shown how well dopamine responses conform to the predictions of this hypothesis, far fewer studies have challenged its implicit assumption that dopamine is not involved in learning value-neutral features of reward. Here, we review studies in rats and humans that put this assumption to the test, and which suggest that dopamine transients provide a much richer signal that incorporates information that goes beyond integrated value.