A Novel fMRI Representation Learning Framework with GAN

Modeling the mapping between mind and brain is the key towards understanding how brain works. More specifically, the question can be formatted as modeling the posterior distribution of the latent psychological state given the observed brain, and the likelihood of brain observation given the latent psychological state. Generative adversarial network (GAN) is known for learning implicitly distributions over data which are hard to model with an explicit likelihood. To utilize GAN for the brain mapping modeling, we propose a novel representation learning framework to explore brain representations of different functions. With a linear regression, the learned representations are interpreted as functional brain networks (FBNs), which characterize the mapping between mind and brain. The proposed framework is evaluated on Human Connectome Project (HCP) task functional MRI (tfMRI) data. This novel framework proves that GAN can learn meaningful representations of tfMRI and promises better understanding of the brain function.