Feedforward extraction of behaviorally significant information by neocortical columns

Area of Science:

• Neuroscience
• Computational Neuroscience
• Computer Vision

Background:

• Neurons in the neocortex are selectively sensitive to sensory input patterns.
• Prevailing theories suggest the cortex prioritizes features with predictable spatial/temporal context.
• Contextually predictable features are crucial for identifying causal environmental factors and have perceptual/behavioral utility.

Purpose of the Study:

• To model the feature extraction process in a prototypical V1 cortical column.
• To investigate how the brain extracts contextually predictable features from sensory input.
• To understand the role of cortical columns in processing perceptually significant information.

Main Methods:

• Implementation of a computational model of a V1 cortical column.
• Training the model on natural images to identify basis features.
• Analyzing the representation of features in upper-layer cortical cells.

Main Results:

• The model identified a limited set of contextually predictable orthogonal basis features in V1 receptive fields.
• Upper-layer cells generated an overcomplete Hadamard-like representation of these features.
• The model demonstrated selective filtering, emphasizing orderly environmental aspects and acquiring higher-order categorization abilities.

Conclusions:

• The proposed feature-extracting circuit captures a core operation of cortical columns in extracting perceptually significant information.
• The model's emergent properties align with known cortical organization, including cell receptive field properties.
• Collective cell participation within a mesocolumn is essential for effective feature tuning.