Challenging the Classical View: Recognition of Identity and Expression as Integrated Processes

Area of Science:

• Cognitive Neuroscience
• Computer Vision
• Machine Learning

Background:

• Classical models propose segregated neural pathways for face identity and facial expression processing.
• Recent neuroimaging suggests common brain regions process both identity and expression information.
• Deep neural networks (DNNs) offer a computational model to investigate these integrated representations.

Purpose of the Study:

• To test if integrated representations of facial identity and expression spontaneously emerge in deep convolutional neural networks (DCNNs).
• To investigate whether DCNNs trained on one task (identity or expression) can generalize to the other.
• To analyze how representational subspaces for identity and expression evolve within DCNNs.

Main Methods:

• Training DCNNs on the CelebA dataset for face identity labeling.
• Training DCNNs on the FER2013 dataset for facial expression labeling.
• Evaluating network performance on the Karolinska Directed Emotional Faces dataset for both identity and expression recognition.
• Employing congruence coefficient analysis to assess feature orthogonality.

Main Results:

• DCNNs trained for identity successfully generalized to expression recognition, and vice versa.
• Networks spontaneously developed representations for both facial identity and expression.
• Features distinguishing identity and expression became increasingly orthogonal across network layers.

Conclusions:

• Deep neural networks spontaneously develop integrated representations for facial identity and expression.
• These findings support the idea that common neural mechanisms may underlie the processing of facial identity and expression.
• DNNs demonstrate a capacity to disentangle representational subspaces for different facial attributes.