Evidence that reward prediction errors account for tonic firing supports a unified model of the dopamine signal

Abstract

It is common to distinguish “stimulus-evoked” spikes from those that appear to occur “spontaneously.” Similarly, the literature on dopamine neurons frequently distinguishes two modes of firing: stimulus-evoked “phasic bursts” that signal “reward prediction error,” and “tonic single spikes” that have been proposed to serve any of a variety of distinct functions. This dualistic account corresponds well with the perspective of an external observer, to whom stimulus-evoked spikes convey information about the stimulus (e.g., reward), whereas ‘spontaneous’ spikes appear uninformative. However, this qualitative distinction may not apply from the perspective of neurons. A unified theory proposes that each spike is caused by greater than expected evidence for reward (“positive reward prediction error”) from the perspective of the dopamine neurons that generate the spikes. Here we analyze firing rates in behaving primates to distinguish unified from dualistic models. The unified model predicts that, in comparison to phasic firing rates, tonic firing rates should exhibit similar but weaker dependence on the reward value of experimenter-controlled stimuli, and that tonic and phasic firing should be positively correlated, both across trials and across neurons. Indeed, both tonic and phasic firing rates increased with reward value, and they were positively correlated across trials and neurons. Dopamine neurons that were more sensitive to phasic reward prediction error exhibited stronger tonic-phasic correlations. Current dualistic models generally lack sufficient detail to predict firing rate statistics, but they imply that to represent distinct variables, tonic and phasic firing should lack statistical dependence. Our results support a unified model in which each spike signals positive reward prediction error.