Duet model of predictive coding unifies diverse neuroscience experimental protocols

The brain continuously generates predictions about the external world, allowing us to rapidly detect and prioritize unexpected events, such as a mistuned piano note or an omitted one. Experimental studies have shown that neurons in sensory cortices respond to various types of contextual deviants across different protocols, and sensory response is not reduced to zero when a stimulus is fully expected. To account for diverse forms of observed deviations, here we introduce duet predictive coding, a minimal and biologically plausible framework in which functional subgroups of neurons encode positive and negative prediction errors separately. In contrast to classical predictive coding [34], which assumes top-down input is purely inhibitory, our theory posits that it is context-dependent rather than absolute. This model reproduces neural responses observed in diverse predictive coding paradigms across visual and auditory cortices. Critically, our framework predicts the existence of neurons tuned to negative prediction errors in the oddball paradigm, a prediction confirmed by our analyses, yet overlooked by classical predictive coding models. Our findings suggest that the brain’s deviance detection relies on duet predictive error computation, offering a unifying explanation across seemingly disparate experimental protocols.