Conference on Cognitive Computational Neuroscience

Despite decades of research, much is still unknown about the computations carried out in the human face processing network. Recently deep networks have been proposed as a computational account of human visual processing. While they provide a good match to neural data throughout visual cortex, they lack interpretability. Here we use a new class of deep generative models, disentangled representation learning models, which learn a low-dimensional latent space that “disentangles” different interpretable dimensions of faces, such as rotation, lighting, or hairstyle. We show that these disentangled networks are a good encoding model for human fMRI data and allow us to investigate how semantically meaningful face features are represented in the brain. We find that several interpretable dimensions, including both identity-specific and identity-invariant dimensions, are distributed widely across the face processing system. The remaining “entangled” representations may be the basis of identity recognition in the brain. These disentangled encoding models provide an exciting alternative to standard “black box” deep learning models and have the potential to change the way we understand face processing in the human brain.