Dissociation and hierarchy of human visual pathways for simultaneously coding facial identity and expression

Area of Science:

• Computational neuroscience
• Cognitive neuroscience
• Machine learning applied to neuroscience

Background:

• Human ability to recognize facial expression and identity simultaneously is effortless but poorly understood.
• Neural mechanisms underlying complex facial recognition tasks require further investigation.

Purpose of the Study:

• To optimize a multi-task deep neural network for simultaneous facial expression and identity classification.
• To elucidate the neural computation underlying human face recognition by comparing a computational model with brain activity.

Main Methods:

• Optimized a multi-task deep neural network for classifying facial expression and identity.
• Analyzed model architecture, layer-specific abilities, and representational similarities.
• Compared model representations with human functional magnetic resonance imaging (fMRI) data from the visual cortex.

Main Results:

• The best-performing model exhibited a 'share-separate' architecture with distinct expression and identity processing branches.
• Specific visual cortex regions (pSTS for expression, aIT/aFFA for identity) correlated with model layers.
• Hierarchical matching observed: aFFA/aIT with high layers, pFFA/OFA with earlier layers.

Conclusions:

• The computational model provides insights into the neural mechanisms of face recognition.
• Human visual system likely employs dissociated and hierarchical neuroanatomical organization for simultaneous facial identity and expression coding.