Organization of face and object recognition in modular neural network models

Area of Science:

• Neuroscience
• Cognitive Science
• Computational Modeling

Background:

• Evidence suggests localized face processing in the brain.
• Double dissociation between prosopagnosia (face recognition deficit) and visual object agnosia indicates partially independent neural mechanisms for faces and objects.

Purpose of the Study:

• To computationally model how face processing specialization arises during development.
• To investigate the roles of competitive selection, infant classification needs, and visual acuity in developing face recognition.

Main Methods:

• Developed two feed-forward computational models of visual processing.
• Utilized a gating network for competitive selection between classification modules.
• Model I: Simple unbiased classifiers; Model II: Biased modules with low/high spatial frequency information.

Main Results:

• Model I demonstrated specialization through competition, where module damage differentially affected face vs. object recognition.
• Model II, with spatial frequency biases, showed stronger face specialization when tasks involved subordinate face classification and superordinate object classification.
• This specialization was not observed with other task/input combinations.

Conclusions:

• Computational models support the theory that face processing specialization can emerge naturally from developmental factors.
• These factors include competitive selection, early need for subordinate face classification, and limited infant visual acuity.
• This suggests a face processing 'module' may not be innately specified but rather a consequence of the developmental environment.