Explaining neural signals in human visual cortex with an associative learning model.

Area of Science:

• Neuroscience
• Cognitive Science
• Computational Neuroscience

Background:

• Predictive coding models propose visual cognition involves matching top-down predictions with bottom-up sensory evidence.
• Neural correlates involve units encoding predictions and units computing prediction errors (surprise).
• This challenges traditional views of neurons as solely bottom-up feature detectors.

Purpose of the Study:

• To formally test predictive coding models using computer simulations.
• To determine if existing neuroimaging data can be explained by associative learning within a neural network framework.
• To investigate neural population signals in the ventral visual stream.

Main Methods:

• Utilized computer simulations of a mathematical neural network model of associative learning.
• Applied the model to analyze human neuroimaging data from a previous study (Egner et al., 2010).
• Compared model outputs (prediction and surprise signals) with neural responses in the fusiform face area (FFA).

Main Results:

• FFA neural population responses were closely fit by model variables representing conditional predictions and their violations.
• The data supported associative prediction and surprise signals over classic feature detection responses.
• Simulations demonstrated that ventral visual stream signals deviating from feature detection can be explained by associative learning.

Conclusions:

• Neural population signals in the ventral visual stream are consistent with predictive coding principles.
• Associative learning plays a crucial role in generating predictions and prediction errors during visual perception.
• The findings support a framework where visual processing involves active prediction and error signaling rather than passive feature detection.