Convergent evolution of face spaces across human face-selective neuronal groups and deep convolutional networks

Area of Science:

• Neuroscience
• Computer Vision
• Artificial Intelligence

Background:

• Deep convolutional neural networks (DCNNs) now achieve human-level performance in complex tasks.
• This advancement provides new avenues for understanding neuronal tuning properties in relation to these tasks.

Purpose of the Study:

• To investigate the relationship between the geometric structure of face representations in biological neural networks and DCNNs.
• To determine if DCNNs with human-level face recognition capabilities share similar face-space geometry with neuronal populations.

Main Methods:

• Recorded intracranial neuronal group activations from 33 patients viewing faces.
• Analyzed pair-wise activation similarities to reveal face-space geometry.
• Compared biological face-space geometry with that of a DCNN trained for face recognition.

Main Results:

• A significant match was found between the face-space geometry of neuronal groups and the DCNN.
• The correlation was strongest in intermediate layers of the DCNN.
• Identity-preserving image manipulations disrupted the DCNN's correlation with neuronal responses.
• DCNN units corresponding to neuronal tuning exhibited viewpoint-selective receptive fields.

Conclusions:

• Convergent evolution of pattern similarities between biological and artificial networks underscores the significance of face-space geometry in face perception.
• The findings suggest human face areas play a role in the pictorial aspects of face perception.
• Face-space geometry is crucial for understanding the visual processing of faces.