Neuroscience is at a moment in history where mapping the connectivity of the human brain non- invasively and in vivo has just begun with many unanswered questions. While the anatomical structures in the brain have been well known for decades, how they are used in combination to form task specific networks has still not been completely explored. Understanding what these networks are, and how they develop, deteriorate, and vary across individuals will provide a range of benefits from disease diagnosis, to understanding the neural basis of creativity, and even in the very long term to brain augmentation. Though machine learning and data mining has made significant inroads into real world practical applications in industry and the sciences, most existing work focuses on lower-level tasks such as predicting labels, clustering and dimension reduction. This requires the practitioner to shoe-horn their more complex tasks, such as network discovery, into the algorithm's settings.
The focus of this grant is a transition to more complex higher-level discovery tasks and in particular, eliciting networks from spatio-temporal data represented as a tensor. Here the spatio-temporal data is an fMRI scan of a person represented as a four dimensional tensor with each entry in the tensor being a data point that indicates the brain activity at that time and location. The overall problem focus is to simplify this data into a cognitive network consisting of identifying active regions of the brains and the interactions that occur between them. The work will consist of three intertwined tasks as follows: i) Supervised and Semi-supervised Network Discovery, ii) Complex Network Discovery and iii) Network Discovery in Populations. In the supervised/semi-supervised setting, the networks discovered involves coordinated activity among some combination of anatomical structures Since all or some of the structures are given along with their boundaries, this is termed a supervised (or semi-supervised) problem. With complex network discovery the team will move beyond finding a single network of coordinated activity to finding multiple networks with complex (beyond coordinates) relationships between the structures/regions. Finally with network discovery in populations , the previous work that studies an individual scan will be expanded to a population of scans. A population may be a collection of individuals performing the same task or a single individual's scans collected over time. Studying such populations allows addressing innovative questions such as: "How does one individual's network change over the course of development, aging, or disease?" and "How do the networks differ for one group of individuals to that of another group?"
