Population coding of reward prediction errors through opponent organization in the fronto parietal network

Computing expectancy violations is essential for decision making and cognitive functions, but its neural mechanisms are incompletely understood. We describe a novel mechanism by which prefrontal and posterior parietal neurons encode reward prediction errors (RPEs) in their population but not single-neuron activity. Simultaneous recordings of neural populations showed that both areas co-activated information about experienced and expected rewards in a precise opponent organization. Neurons encoding expected rewards with positive (negative) scaling were reactivated simultaneously with those encoding experienced rewards with negative (positive) scaling. This opponent organization was mirrored in polarity-dependent noise correlations. Moreover, it extended to two types of expectancy information – based on task-relevant visual cues and statistically irrelevant reward history - allowing decoding of signed and unsigned RPE in two reference frames. Frontal and parietal areas implement canonical computations that facilitate contextual comparisons and the readout of multiple types of expectancy violations to flexibly serve behavioral goals.