Top-down perceptual inference shaping the activity of early visual cortex.

Area of Science:

• Computational neuroscience
• Machine learning
• Systems neuroscience

Background:

• Deep discriminative models offer insights into brain's hierarchical processing but differ computationally from biological systems.
• These models require supervised learning signals and rely on feed-forward processing, unlike the brain's top-down connections.

Purpose of the Study:

• To develop a hierarchical deep generative model that addresses limitations of current models.
• To investigate how hierarchical representations and top-down influences shape neural activity in the visual cortex.

Main Methods:

• Development of a hierarchical deep generative model.
• Testing the model's ability to predict experimental results in primary (V1) and secondary (V2) visual cortices.
• Analyzing the model's predictions regarding texture sensitivity and top-down interactions.

Main Results:

• The model successfully predicts a wide range of experimental findings in V1 and V2.
• Neural sensitivity to textures in V2 is explained as a result of learning hierarchical image representations.
• Top-down influences are shown to be intrinsic to hierarchical inference, impacting V1 responses.

Conclusions:

• Hierarchical deep generative models offer a more biologically plausible alternative to discriminative models.
• The model elucidates the neural basis of texture perception and top-down modulation in the visual system.
• Hierarchical inference provides a framework for understanding how higher-level representations influence lower-level neural activity.