Functional magnetic resonance imaging is a technology allowing for a non-invasive measurement of the brain activity. Data are encoded as sequences of 3D images, usually few hundreds samples, each made by tens of thousands voxels, namely volumetric pixels. The main question in neuroimaging is the identification of the voxels affected by a specific brain activity. This task, referred to as brain mapping, can be conceived as a problem of feature rating. The challenge is twofold: the former is to deal with the high feature space dimensionality; the latter is the need for preservation of redundant features. Most common techniques of feature selection do not cover both requirements. In this work we propose the adoption of a random subspace method, arguing, by theoretical arguments and empirical evidence on synthetic data, that it might be a viable solution for a multi-variate approach to brain mapping. In addition we provide some results on a neuroscientific case study investigating on a visual perception task.
展开▼
