Explaining face representation in the primate brain using different computational models.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Computer Vision

Background:

• Understanding neural object representation is key in visual neuroscience.
• The macaque face patch system in inferotemporal (IT) cortex is crucial for face processing.
• Previous work suggested face neurons encode axes of an active appearance model.

Purpose of the Study:

• To systematically compare computational models, including deep neural networks, in explaining neural responses in the anterior medial (AM) face patch.
• To investigate how well different models capture neural representations of face identity.

Main Methods:

• Recorded neural responses from AM face patch cells using a large dataset of real face images.
• Compared the explanatory power of various computational models, including the active appearance model and convolutional neural networks (CNNs).

Main Results:

• The active appearance model explained neural responses well, but CORnet-Z, a general object classification CNN, performed comparably or better.
• Deep neural networks specifically trained for facial identification showed poorer performance in explaining neural responses.
• CNN units were less sensitive to non-identity factors like illumination compared to neurons.

Conclusions:

• Feedforward deep neural networks trained on general object recognition may offer better models of neural object representation than specialized face models.
• Neural object representation in the AM face patch is influenced by factors beyond core identity, which are not fully captured by current specialized deep learning models.