Deep learning detects subtle facial expressions in a multilevel society primate

Area of Science:

• Primate ethology
• Artificial Intelligence in Biology
• Animal Communication

Background:

• Facial expressions in nonhuman primates are complex, involving psychological, emotional, and physiological factors.
• The functional significance of subtle primate facial expressions remains uncertain.
• Golden snub-nosed monkeys (Rhinopithecus roxellana) live in complex, multilevel societies.

Purpose of the Study:

• To investigate the existence and detectability of subtle facial expressions in golden snub-nosed monkeys.
• To compare the accuracy of artificial intelligence (AI) models versus human observers in recognizing these expressions.
• To explore the role of subtle facial expressions in primate social systems.

Main Methods:

• Collected 3427 front-facing images of golden snub-nosed monkeys from video clips.
• Utilized three deep learning models (EfficientNet, RepMLP, Tokens-To-Token ViT) for AI-driven facial expression recognition.
• Compared AI accuracy with human observer accuracy (inexperienced and experienced groups).

Main Results:

• Human observers detected facial expressions with an average accuracy of 32.1% (inexperienced) and 45.0% (experienced), near chance level (33%).
• AI deep learning models achieved significantly higher accuracy rates.
• The best-performing AI model reached an accuracy of 94.5% in detecting subtle facial expressions.

Conclusions:

• Golden snub-nosed monkeys exhibit subtle facial expressions undetectable by human observers.
• AI is a powerful tool for uncovering complex communication signals in animals.
• Subtle facial expressions likely play a crucial role in the evolution of complex primate social systems.