A wide variety of brain-derived signals presently are available to drive brain computer interface devices. These include the popular EEG recordings, magnetoencephalography, functional magnetic resonance imaging (fMRI), near-infrared spectroscopy and others. Each are known to be quantitatively altered by intentional mental activity and, with the power provided by statistical machine learning, each to varying degrees may be decoded to the end of controlling devices. I will speak to the question of the reverse challenge of understanding better the operations of the brain through analysis of the control signals and argue that careful selection of the features themselves might serve the dual purposes of improving the efficiency and accuracy of the brain-computer interface, and serve to improve our understanding of the underlying neurophysiology. The discussion will focus on the use of brain network features exposed through fMRI and on understanding the temporal dynamics of the individual features and their state transitions.