Disentangled deep generative models reveal coding principles of the human face processing network.

Area of Science:

• Neuroscience
• Computer Science
• Cognitive Science

Background:

• The human face processing network's computations remain largely unknown.
• Deep neural networks model visual processing but lack interpretability.
• Interpreting brain activity is crucial for understanding face recognition.

Purpose of the Study:

• To develop an interpretable computational model for human face processing.
• To investigate the representation of facial features in the brain using deep learning.
• To compare disentangled representation learning with standard deep learning approaches.

Main Methods:

• Utilized disentangled representation learning models to create a low-dimensional latent space for faces.
• Enforced statistical independence between latent dimensions in an unsupervised manner.
• Assessed interpretability of latent dimensions using human raters and modeled fMRI data.

Main Results:

• Most learned latent dimensions were interpretable, representing semantic face variations (e.g., rotation, lighting).
• These dimensions effectively encoded human functional magnetic resonance imaging (fMRI) data.
• Found a topographical organization of face features in face-selective regions, with identity-relevant and irrelevant features distributed across the network.

Conclusions:

• Disentangled face encoding models offer a powerful, interpretable alternative to black-box deep learning for neuroscience.
• This method advances our understanding of how the brain represents and processes facial information.
• Revealed insights into the neural basis of facial identity representation and feature segregation.