Cortical circuit implementation of signed deviant detection with positive and negative prediction error neurons

Predicting future events is fundamental for animals to adaptively interact with their environment, but whether and how the cerebral cortex processes predictive information remains an unresolved question. We combined in vivo two-photon calcium imaging in mouse somatosensory cortex with an oddball paradigm that decouples sensory stimulus properties from predictability context. We recorded neuronal responses in layer (L)2/3, bottom-up sensory inputs, and top-down prefrontal feedback. We find that L2/3 pyramidal neurons concurrently encode stimulus features and statistical context. We identify coexisting positive prediction error (PE) neurons that subtract predictions from incoming sensory inputs and negative PE neurons that do the opposite. This architecture represents a unified computational principle where error signals scale proportionally with the magnitude of mismatch across distinct stimulus features tested (amplitude, direction, omission). Furthermore, bottom-up pathways operate as unsigned novelty detectors, L2/3 circuit computes signed PE, and top-down projections signal a precision-weighted signed error, demonstrating hierarchical predictive processing.