Improved brain pattern recovery through ranking approaches

The prediction of behavioral information or cognitive states from brain activation images such as those obtained with fMRI can be used to assess the specificity of several brain regions for certain cognitive or perceptual functions. This kind of analysis is implemented by learning a classifier or regression function that fits a given target variable given fMRI activations. The accuracy of this prediction depends on whether it uses the relevant variables i.e. the correct brain regions. Recovering the truly predictive pattern has proven to be challenging from a statistical point of view: the high dimensionality of the data together with the limited number of images makes the problem of brain pattern recovery an ill-posed problem. So far, the approaches proposed to address this issue have relied on linear models, with univariate, i.e. voxel-based, Anova (analysis of variance) for hypothesis testing, or, for predictive modeling, with the choice of a regularizer using a priori domain-specific knowledge, such as the l