Emerged human-like facial expression representation in a deep convolutional neural network.

Area of Science:

• Cognitive Science
• Computer Vision
• Neuroscience

Background:

• Deep convolutional neural networks (DCNNs) trained for facial identity recognition can spontaneously learn features for facial expression recognition.
• The precise role of domain-specific versus domain-general experience in developing these abilities remains unclear.

Purpose of the Study:

• To investigate whether expression-selective units in DCNNs exhibit human-like expression perception characteristics.
• To determine if face-specific experience or domain-general processing is necessary for the emergence of human-like facial expression perception in DCNNs.

Main Methods:

• Analysis of a VGG-Face model trained for facial identification to identify expression-selective units.
• Comparison of expression-selective units in a VGG-16 model trained for object classification and an untrained VGG-Face.
• Assessment of human-like hallmarks such as facial expression confusion and categorical perception in these units.

Main Results:

• Expression-selective units emerged in a VGG-Face trained for facial identification, showing tuning to basic expressions and human-like perception characteristics.
• Similar expression-selective units were found in DCNNs trained on object classification and untrained models, but they lacked human-like perception hallmarks.
• These findings indicate that domain-general processing alone is insufficient for developing human-like facial expression perception.

Conclusions:

• Domain-specific visual experience with face identity is essential for developing human-like facial expression perception in DCNNs.
• The findings underscore the significant contribution of nurture and specific learning experiences in shaping perception.
• This research provides insights into the computational mechanisms underlying human facial expression recognition.