Strategies for Discovering Mechanisms of Mind using fMRI Joseph Ramsey & Clark Glymour
Abstract: Functional Magnetic Resonance Imaging (fMRI) is a useful domain for discovering causal mechanisms in the brain; samples are sufficiently independent to allow the application of a number of techniques that in their simple forms assume i.i.d. sampling. A number of such search procedures are possible, for single-subject or multi-subject analysis, including PC, GES, IMaGES, GIMME. These are capable of inferring undirected causal connections with a fair degree of accuracy, as judged from the best available simulation studies of fMRI activity in the brain. Inferring orientations is dicier, although fMRI signals are sufficiently non-Gaussian to apply a variety of techniques that orient known causal connections by appeal to higher moments of the signals. All of this work has been done with relatively small graphs, often less than 5 regions of interest (ROIs), with the maximum in the simulation literature of 50 nodes. We propose to scale this number up and find causal connections and orientations for thousands of nodes, approaching the number of voxels in the cortex. We set out some of the difficulties with this approach and show some successful work already accomplished in this direction.