Machine Learning Opportunities in Analyzing Brain Imaging Data

Francisco Pereira

Contemporary Cognitive Neuroscience relies to a large extent on the use of functional MRI (fMRI) scanners. As with other modern scientific instruments, the output of a fMRI scanner goes beyond what can be viewed with the naked eye or fathomed manually. fMRI datasets are large, must be heavily processed before use and conclusions can only be safely drawn through statistical methods.

In this talk I'll introduce fMRI and how the most common experimental question (essentially "which brain areas light up when you do X?") is answered using generalized linear models. I'll contrast this with the new experimental questions that have been posed and answered using machine learning classifiers and give a brief review of that work. Finally, I will present a few case studies where domain expert questions led to either the adaptation or the development of new machine learning methods to answer them; my aim is to illustrate how machine learning can both influence scientific practice and derive inspiration from having to deal with the characteristics and research goals in this domain.