Classication and discrimination of functional Magnetic Resonance Image (fMRI) scans using machine learning is a eld of interest to the medical community because of its potential to provide an automated way of detecting neurological conditions while simultaneously giving insight into potential causes of these disorders. Discrimination has been successfully executed using spatially localized signal dierences to distinguish among groups of patients [FFM03] [ZS05], but is limited by such characteristics as a need for spatial alignment among fMRI scans and its dependence on local, rather than global, activation dierences. In this dissertation, we present methods of classifying and discriminating among fMRI scans using a network-driven view of activity rather than a local-driven approach. By modeling the fMRI scans as linear combinations of independent components, we demonstrate how these potential basis functions serve as a concise encoding of activity. This encoding is suitable for classication among patient groups such as Schizophrenic/Control, Alzheimer's/Old/Young, and IBS/Control. Methods are outlined for distinguishing among behavior based upon component activity differences within a subject, classifying patient groups using unaligned fMRIscans, classifying patient groups using aligned scans and identifying common deviant networks, identifying treatment eects between scan sessions while holding a task constant, and methods of modeling scans using the independent components as basis functions. This collection of methods points to the eectiveness of using sparse encoding models such a independent components analysis applied to medical imaging as a dimension reduction and signal extraction technique.