Predicting Behavior from Functional Magnetic Resonance Imaging

A canonical problem in machine learning seeks to predict an outcome based on a set of features. When applied to fMRI data, this problem can be viewed as prediction of behaviour (the outcome) based on a set of features (approximately 10^6 space-time voxels for each outcome). A common approach to such problems is one of penalized regression using a penalty such as the LASSO to automatically do feature detection. In the imaging context, some level of smoothness is also a desirable property, a property which the LASSO lacks. We discuss some generalizations of the LASSO that yield sparse structured regions which can be used to predict subjects’ behavior. The mathematics of such methods are intimately related to those of compressive sensing, which will be discussed by Prof. Lustig. They are also related to the geometry discussed by Prof. Adler. The talk will describe the prediction problem and the resulting prediction model, as well as some of the geometrical connections this problem has with signal detection and image acquisition.