Brain as a self-predictor: Sparse full-brain auto-regressive modeling in fMRI

We demonstrate a method to build an autoregressive model for the whole brain without carrying out any aggregation of the fMRI data. The model gives biologically meaningful results and has several desirable properties. We show that the model gives significantly improved prediction on unseen data as compared to baseline methods. The voxels with better prediction are distributed throughout the brain, including the task positive and task negative regions. In addition to the active regions identified by the general linear model (GLM), our analysis also uncovers complex interactions among the regions involved in the default mode networks.