Unsupervised inference approach to facial attractiveness.

Area of Science:

• Cognitive Science
• Computer Vision
• Machine Learning

Background:

• Facial attractiveness perception is complex, involving individual features and their interplay.
• Current machine learning models often use subject-averaged ratings, potentially oversimplifying the phenomenon.
• Recent studies show individuals can modify faces in a 'face-space', indicating subjective preferences.

Purpose of the Study:

• To investigate the set of sculpted facial vectors in subjective attractiveness experiments.
• To infer minimal, interpretable, and accurate probabilistic models of preferred facial variations.
• To explore the relationship between facial preferences and observer characteristics, such as gender.

Main Methods:

• Unsupervised inference of probabilistic models (Maximum Entropy, ANNs) for sculpted facial variations.
• Generative models were applied to supervised classification tasks.
• Analysis of classification accuracy based on the order of non-linear interactions between facial landmarks.

Main Results:

• Probabilistic models successfully captured inter-subject variance in facial sculpting.
• Gender of the sculptor could be predicted with high accuracy from sculpted facial vectors.
• Classification accuracy increased with higher-order interactions among facial landmarks.

Conclusions:

• Cognitive mechanisms for facial discrimination likely involve interactions between multiple facial landmarks, not just individual positions.
• Subjective facial preferences may encode significant information about observer characteristics, aligning with the multiple motive theory of attractiveness.
• Machine learning models incorporating higher-order landmark interactions may better capture the complexity of attractiveness perception.