Bayesian Models for Combining Data Across Domains and Domain Types in Predictive fMRI Data Analysis

In the context of predictive fMRI data analysis, in particular classification-based analysis, the analysis is usually done separately for a particular subject in a specific study, or by combining data across subjects and studies after normalization to a common template by assuming that there are no inter-subject and inter-study variations. These approaches are suboptimal for a number of reasons. Based on findings in the cognitive neuroscience field, there is a reason to believe that data from other subjects and from different but similar studies exhibit similar patterns of activations, implying that there is some potential for leveraging data from other subjects. However, each subject's brain might still exhibit some variations in the activations compared to other subjects' brains, based on factors such as differences in anatomy, experience, or environment. Furthermore, current normalization techniques might blur and distort the data, causing some loss of information.

I propose to investigate and develop principled Bayesian methods to tackle these problems, and by doing so, enable combining data across domains and domain types, of which subjects and studies are instances. One goal is to improve predictive performance when compared with predictive analysis of each domain and domain type separately, especially in cases where there is a lot of commonality of activations across different domains and domain types. Another goal is to figure out the extent of commonality of various cognitive phenomena across domains and domain types.

Thesis Committee:

Tom M. Mitchell, Chair
Zoubin Ghahramani
Eric Xing
David Blei, Princeton University

Proposal (slides), April 3, 2007, 1.30pm, Wean 8220