Abstract

The advent of functional Magnetic Resonance Imaging (fMRI) has provided researchers with a method of probing the activity of the brain with fine granularity. One goal of fMRI research is to use brain activity to classify the task a subject is performing. Due to the vast quantity of data in fMRI studies, feature selection techniques are used to select relevant brain areas (features) as input to a classifier. In this problem setting, features can be selected through two methods:(1) discriminability tests that compare the feature’s activity between the specific tasks to be classified or (2) activity tests that compare the task activity against activity during a rest condition.
A surprising result is that features chosen using activity tests often provide better classification accuracy than discriminability tests, despite receiving no information pertinent to the classification task. The goal of my research has been to use a plausible model of fMRI data coupled with mathematical analysis and experiments using synthetic data to explain this phenomenon. Realistic parameters to this model allow very specific predictions about the performance of different feature selection methods. Subsequently, this approach is validated and extended through the investigation of experimental data in a semantic categories experiment. Finally, the results of this approach are used to explore alternative feature selection methods that have the potential to outperform activity-based feature selection. A case study comparing feature selection methods in experimental data serves to underscore the applications of this research. Although this research examines fMRI data, the research methodology, analytical …

Date

January 1, 1970

Authors

Jay Pujara

Institution

Carnegie Mellon University Pittsburgh, PA