Understanding Human Cognitive Control via fMRI Analysis

Cognitive control is the essential high-order information processing system of human brains. Understanding cognitive control helps improve the diagnostic and treatment of various neurological disorders. We focus on learning the uncertainty representation in cognitive control, namely, how human brains react to the same task with different levels of uncertainty, using task-evoked function MRI images. The learning includes two tasks: identification of key brain regions and brain connectivities. We propose an interpretable convolutional neural network, called ROI-reweight 3D CNN, to identify key brain regions. We train a classifier for task-evoked fMRI images, which also locates crucial ROIs based on a reweight layer. Brain connectivity analysis can be formulated as a graph inference problem, in which the edges in the graph indicate relations between ROIs. We propose a neural architecture based on Markov Random Fields (MRF) for the brain network learning task. The neural network learns a graphical model as brain connectivity pattern. Furthermore, we are interested in learning the differential connectivity patterns under different uncertainty conditions. We design a neural network architecture which learns to decide whether two input images are from the same class (uncertainty level). The key is to identify an underlying graphical model (MRF) that captures the difference between different uncertainty levels