Hierarchical substrates of prediction in visual cortical spiking

Predictive processing models have recently flourished in neuroscience^1-9. Feedforward and feedback modulation are at the heart of these hierarchical predictive processing models. Previous experimental studies using fMRI, EEG/MEG, and LFP^1,10,11 could not reliably resolve feedback modulation from local computations and feedforward outputs. Here, using open-science⁹, multi-species, multi-area, high-density¹², laminar neurophysiology¹³, we empirically test whether predictive processing is a key component shaping sensation. To isolate sensory information processing and eliminate motor/reward confounders^1,10,11, we use a no-report task. Our task leveraged so-called global oddballs (GO) as unpredictable, deviant stimuli that circumvent low-level adaptation. We examined their responses relative to local oddballs (LO) that we habituated into highly predictable priors. Four surprising findings in this dataset challenge many existing predictive processing models. First, passively evoked GO responses were exclusive to higher-order, more cognitive areas rather than early-to-mid sensory cortex. Second, interneuron-targeted optogenetics revealed no evidence for a subtractive mechanism in both primates and mice. Third, highly predictable LO responses dominated in over 50% of all neurons, including in higher-order cortex which should have anticipated them, indicating limited evidence for predictive suppression. Lastly, prediction errors followed a feedback, rather than a feedforward signature. These results reveal circuit dynamics that govern the shaping of sensory processing by prediction, which will motivate new, neurally-constrained predictive processing models.