Experience Alone Can Generate Human Face Specialization: Evidence From Deep Learning Models

Area of Science:

• Cognitive Science
• Artificial Intelligence
• Computer Vision

Background:

• Human face recognition shows specialization for upright, own-race faces.
• The causal role of experience in this specialization is hard to study in humans due to uncontrolled natural exposure.
• Deep learning models exhibit human-like face recognition biases, offering a controllable system to study experience.

Purpose of the Study:

• To investigate the sole contribution of experience to human-like face specialization in artificial systems.
• To examine how varying amounts of face experience affect face inversion, other-race, and other-age effects in deep neural networks.

Main Methods:

• Systematically manipulated the quantity of face experience during deep neural network training.
• Assessed the impact of training experience on recognition performance for upright/inverted and own-group/other-group faces.

Main Results:

• Increased own-group upright face experience amplified other-group and face inversion effects.
• Performance gains were concentrated on upright, own-group faces, with minimal improvement for other-group or inverted faces.
• Deep neural networks demonstrated human-like face specialization driven solely by training experience.

Conclusions:

• Experience alone is sufficient to generate human-like face specialization effects in artificial systems.
• Selective improvement in recognizing upright, own-group faces is a direct consequence of increased exposure to this category.
• Deep learning models provide a valuable tool for understanding the mechanisms of human perceptual expertise.