Constrained inference in sparse coding reproduces contextual effects and predicts laminar neural dynamics.

Area of Science:

• Computational Neuroscience
• Systems Neuroscience
• Visual Perception

Background:

• Neurons in the primary visual cortex (V1) exhibit complex responses to stimuli outside their classical receptive field.
• Existing sparse coding models explain some contextual effects but lack realistic neural structures and mechanisms.

Purpose of the Study:

• To propose an extended generative model for visual scenes incorporating spatial feature dependencies.
• To derive a neurophysiologically plausible inference scheme for V1 processing under local information constraints.

Main Methods:

• Developed an extended generative model for visual scenes with spatial dependencies.
• Derived a realistic neural inference scheme constrained by local image information access.
• Simulated the model using natural images and contextual stimuli.

Main Results:

• The model predicts a connectivity pattern in V1 matching observed horizontal and feedback projections.
• It replicates hallmark contextual processing effects, including surround modulation differences between neural populations.
• The model's structure involves two V1 neural populations in orientation hypercolumns with recurrent interactions.

Conclusions:

• The proposed model offers a novel framework for understanding contextual processing in the visual system.
• It assigns a functional role to horizontal axons and feedback projections in V1.
• The model provides a neurophysiologically realistic explanation for complex V1 responses to contextual stimuli.