A shared predictive architecture in the sensory cortex for statistical and reward-based learning

The sensory cortex is widely considered a feed-forward feature extractor whose stimulus-driven responses can be modified by context or experience. Recent work, however, suggests a richer role. In implicit, statistical learning, mismatch responses in primary auditory and visual areas point to a computation of sensory prediction errors. In explicit, reward-based learning, converging evidence shows that populations in the sensory cortex rapidly develop reward prediction activity, firing when an expected outcome would follow the stimulus. These results demonstrate that, alongside feature representation, the sensory cortex has another core function: prediction. This predictive role could, in principle, be implemented by one circuit motif: feed-forward drive to basal dendrites, distal input to apical tufts, and local disinhibition. When the distal input carries a sensory prediction, the sensory cortex computes prediction errors by directly comparing predicted (distal) with actual, feedforward drive. When it carries outcome information, the sensory cortex computes a prediction, without computing errors. We assemble the empirical foundation for this possibility drawn primarily from the auditory cortex, complemented by convergent results from other sensory cortices and then provide a roadmap to test this principle.