Machine Learning Approaches to fMRI
Tom Mitchell
Carnegie Mellon University
AI and Learning
Machine learning algorithms, such as classifier learning methods, are increasingly used to analyze fMRI data. Whereas many traditional approaches to fMRI analysis seek to answer the question, “Which brain regions are active on average during cognitive task X?”, classifier learning methods address questions such as, “Is the subject performing cognitive task X or Y during this particular time interval?” The cross-validated accuracy of classifier predictions provides an objective, assumption-free basis for assessing the validity of learned models.
This lecture covers classifier learning methods such as Support Vector Machines (SVMs) and Bayesian classifiers, and their application to fMRI data analysis. Several examples are discussed, including training classifiers to distinguish whether a human subject is comprehending a sentence or a picture, and whether the subject is reading a word related to tools or buildings. Trained classifiers for these tasks achieve prediction accuracies ranging from 80% to 95% on new data from the same subject, and in some cases can generalize to new subjects outside the training set.
We examine feature selection methods that are crucial to achieving these accuracies and consider the role of these classifiers in studying human cognitive processing. Related time series modeling approaches based on temporal Bayesian networks are also explored, along with their potential role in modeling cognitive processes.
