High-level visual prediction errors in early visual cortex.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Cognitive Science

Background:

• Perception integrates sensory input with prior knowledge.
• Predictive processing models suggest neural activity increases with surprising stimuli.
• The hierarchical origin of prediction errors remains unclear.

Purpose of the Study:

• To investigate whether prediction errors arise locally or are broadcast from higher cortical levels.
• To determine the nature of surprise signals in the cortical hierarchy.
• To test hypotheses using functional magnetic resonance imaging (fMRI) and deep neural network (DNN) models.

Main Methods:

• Utilized fMRI in human volunteers to measure neural activity.
• Employed deep neural network (DNN) models to simulate visual processing.
• Compared neural responses to stimuli varying in low-level and high-level visual surprise.

Main Results:

• Neural activity in both low- and high-level visual cortex correlated with high-level surprise.
• Responses in early visual cortex scaled with high-level surprise, deviating from feedforward predictions.
• No significant response to low-level visual surprise was observed in early visual areas.

Conclusions:

• Prediction errors are computed at higher cortical levels and broadcast to earlier areas.
• High-level predictions exert top-down influence on sensory processing in early visual cortex.
• This top-down modulation aids perceptual inference by constraining sensory information.