Face identity coding in the deep neural network and primate brain.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Artificial Intelligence

Background:

• Understanding neural encoding of face identities is a key challenge in face perception.
• Existing models lack the capacity to analyze the entire face processing network and numerous facial features.

Purpose of the Study:

• To investigate how neurons encode face identities using in silico experiments with deep neural networks (DNNs).
• To compare the representational capacity of artificial neural units with primate neural recordings.

Main Methods:

• Conducted in silico experiments utilizing a pre-trained face recognition DNN with diverse stimuli.
• Identified identity-selective units within the DNN and analyzed their discriminability to novel faces.
• Compared DNN unit responses to single-neuron recordings from monkey and human primates using identical stimuli.

Main Results:

• Identified DNN units selective for face identities, demonstrating generalized discriminability to new faces.
• Visualizations and manipulations confirmed the critical role of these identity-selective units in face recognition.
• Found that artificial units in DNNs share similar facial feature representations and region-based coding mechanisms with primate neurons.

Conclusions:

• DNNs provide a valuable model for understanding primate face recognition.
• Direct comparison between artificial and primate neural systems illuminates the mechanisms of face identity encoding in the brain.
• The study bridges computational models and biological neural systems to explain face perception.