Learning Common Features from fMRI Data of Multiple Subjects

Functional Magnetic Resonance Imaging (fMRI), a brain imaging technique, has allowed psychologists to identify what parts of the brain are involved in various tasks. Recently, Mitchell et al (2003) have used fMRI in a novel way: to infer from it a person’s mental states using machine learning algorithms. Wang, Hutchinson, and Mitchell (2003) have extended these algorithms to make predictions across subjects, using hardcoded common representations. We have gone further to develop cross-subject clustering, a method of learning common representations. This method not only offers the theoretical advantage of learning, but also appears to offer the empirical advantage of improved cross-subject predictions. The empirical studies were limited to a single dataset (Sentence-then-Picture), however, so further work is needed to confirm its general utility. Several of our other experiments demonstrate that unsupervised learning generally attains accuracies nearly as high as those of supervised learning across subjects, and in some cases higher. Finally, we briefly catalog several other less successful approaches to the cross-subject prediction problem. Outline