Deep models of superficial face judgments.

Area of Science:

• Computer Vision
• Cognitive Science
• Psychology

Background:

• Human faces are diverse, leading to varied inferences of psychological traits and attributes.
• Deep neural networks offer rich face representations but lack interpretability.
• Understanding the link between face features and human perception is challenging.

Purpose of the Study:

• To model human inferences of over 30 attributes from faces.
• To explore the relationship between high-dimensional face representations and perceived attributes.
• To create a predictive model for face-based attribute inference.

Main Methods:

• Utilized deep generative image models.
• Incorporated over 1 million human judgments of face stimuli.
• Modeled inferences across a comprehensive latent face space.
• Combined machine learning with large-scale behavioral data.

Main Results:

• Achieved predictive accuracy approaching human interrater reliability.
• Demonstrated the necessity of large datasets and high-dimensional representations.
• Successfully modeled inferences for more than 30 distinct attributes.
• Validated the model's ability to predict and manipulate perceived attributes.

Conclusions:

• The developed model provides a robust framework for understanding face perception.
• Deep generative models can effectively bridge the gap between computational representations and human inferences.
• The model has applications in predicting, manipulating, and generating face stimuli with specific perceived attributes.