Understanding how our brain performs cognition and perception is a challenging endeavor. Recently, machine learning approaches have emerged as an interesting direction to study our cognitive architecture based on imaging techniques such as functional magnetic resonance imaging (fMRI). In this talk I will discuss two approaches that aim to capture and describe functional interaction patterns that emerge during specific cognitive processes and can span the entire brain.
In the first part we explore if the predictive power of fMRI signals with regard to an experimental condition (such as different objects shown to a person) can serve as a marker of specific activation patterns of functional processes involving multiple regions in the brain. The second part of the talk will focus on the exploratory analysis of the interaction patterns in fMRI data. We use a diffusion process on the set of fMRI voxels and a corresponding spectral embedding to establish a distance that captures the functional relationship between different cerebral regions. The resulting geometry captures the global interaction pattern within the brain. It makes the exploration of the entirety of functional interactions, and the mutual roles of individual regions possible. Furthermore, initial results indicate that this map is a stable descriptor of the brain state, and allows to decouple cognitive processes and their anatomical embedding. To illustrate the approaches and highlight the open questions I will discuss results in a study of migration mechanisms affecting language areas caused by tumor growth; a comparison of functional architecture involved in reward recognition between control subjects and cocaine abusers; and the structure of high-level human object category recognition in the visual pathway.