Discovering brain activity patterns associated with specific stimuli, behaviors, or mental disorders is a primary objective of many neuroimaging studies. In particular, functional magnetic resonance imaging analyses are often concerned with finding brain areas “most relevant” to a mental state of interest. In this paper, we consider the predictive accuracy of a subset of voxels (smallest units of a three-dimensional image) identified by multivariate sparse regression as a better relevance measure than the commonly used mass-univariate correlations between the task and individual voxels. Sparse regression detects synergistic multivoxel interactions missed by univariate techniques. Moreover, when invoked iteratively, it allows us to discover novel patterns of information distribution through the brain. Indeed, unlike univariate techniques that typically imply a sharp transition from task-relevant to task-irrelevant brain areas, our approach reveals that for some tasks, task-relevant information is distributed rather “holographically” through the entire brain, yielding multiple relevant subsets of voxels, with gradually declining predictive accuracy, and without any clear separation between task-relevant and task-irrelevant ones. Interestingly, as our preliminary results show, such phenomena occur in certain “complex” experiences such as pain perception, but not in relatively “simple” tasks such as finger tapping.