Machine learning approaches towards holistic brain functional space discovery from fMRI big data

While functional neuroimaging has been going through significant advancement in the past decade, there remains a fundamental question of how we can utilize the imaging data to describe brain functional behavior in a reproducible and faithful manner. In this dissertation I have elucidated a series of my works all aims at answering the above question, yet from three different perspectives both regarding to the neuroscience implication and to the scale of the data. Firstly, statistical models are built to characterize the changes of functional organization pattern in individual brains (small size), in order to detect the quasi-stable brain states. Secondly, the concept and corresponding framework of functional connectomics summarize the common connectivity patterns within a group of individuals (medium size), and use them for dynamic transition modeling. Thirdly, dictionary learning method and its distributed implementation enable us for the efficient functional network discovery from population-wise data (large size). Based on these works, we could then eventually learn the set of holistic brain functional space from fMRI big data, through which individual signals can be effectively encoded and analyzed.