ABSTRACT Alzheimer’s disease and associated dementias are major public health challenges with a multifold increase in prevalence expected in the coming decades. Alzheimer’s disease is increasingly recognized as having network- level effects and interactions. In this project, we will develop a deep learning model to learn the latent representation of functional neuroimaging, in order to disentangle the underlying sources and better reconstruct the data. Deep learning approaches in fMRI have faced a common challenge on generalizability and explainability. To address these issues, the system will learn representations that can be decoded and interpreted as spatial patterns and temporal dynamics of brain networks; and be readily generalizable to different subjects, brain states, behavioral tasks, and disease conditions without a need to redesign or retrain the system from scratch. The proposed focus on Alzheimer’s disease is the first step in exploiting this notion for clinical application. We will leverage both publicly available large data (e.g., Human Connectome Project-Aging, Alzheimer’s Disease Neuroimaging Initiative) as well as the well-characterized longitudinal cohort of the NIA P30-funded Michigan Alzheimer’s Disease Research Center (MADRC); this cohort undergoes annual neurological and neuropsychological evaluations and is particularly unique since it consists of ~45% African Americans. This research is particularly relevant for ethnic minority populations since African Americans are almost twice as likely to develop cognitive decline as Non-Hispanic white Americans; yet most of what has been learned about dementia biomarkers is based on study samples that are primarily Non-Hispanic white Americans. The overall goal of this project is to develop an enhanced deep learning model for improved data representation, subtype classification and prediction of clinical behavioral measures and apply it to the domain of mild cognitive impairment (MCI) and dementia of the Alzheimer’s Type (DAT).