Stimulus history, not expectation, drives sensory prediction errors in mammalian cortex

Predictive coding (PC) hypothesizes that the brain computes internal models of predicted events and that unpredicted stimuli are signaled with prediction errors that feed forward. We tested this hypothesis using a visual oddball task. A repetitive sequence interrupted by a novel stimulus is a "local" oddball. "Global" oddballs defy predictions while repeating the local context, thereby dissociating genuine prediction errors from adaptation-related responses. We recorded neuronal spiking activity across the visual hierarchy in mice and monkeys viewing these oddballs. Local oddball responses largely followed PC: they were robust, emerged early in layers 2/3, and fed forward. Global oddball responses challenged PC: they were weak, absent in most visual areas, more robust in prefrontal cortex, emerged in non-granular layers, and did not involve inhibitory interneurons relaying predictive suppression. Contrary to PC, genuine predictive coding does not emerge early in sensory processing, and is instead exclusive to more cognitive, higher-order areas. Core principles of Predictive Coding explain local, but not global oddball brain responsesGlobal oddballs are sparsely encoded in visual cortex and do not engage feedforward mechanismsParvalbumin and somatostatin interneurons do not support sensory global oddball detectionGlobal oddballs are earliest, and most robustly encoded, in prefrontal brain areas. Global oddballs defy our expectations without standing out from their local context. Whether they are encoded in sensory brain areas was uncertain. Through multiarea neurophysiological recordings in mice and monkeys, Westerberg et al. show that global oddballs neuronal responses contradicts predictive coding hypotheses and emerge in higher order areas.