Mining brain images to uncover cognition and neuropathologies.
How do thoughts, sensations, decisions, emotions, arise from neural activity? What is the biology underlying a social psychiatric disorder such as autism? These questions are answered in cognitive neuroscience via experimental psychology. Unlike physics, cognitive psychology does not rely on first principles. However, brain imaging provides quantitative measurements that can build or constrain model of brain function.
I will discuss how we use build machine learning tools to help understanding the link between the brain, the organ, and the mind, our mental world. Building on the key concepts of machine learning, seen as high-dimensional statistics, I will present progress on "brain reading", inferring thoughts from brain signals, and conversely, finding the characteristics of stimuli salient for the brain.
To study brain pathologies, experiments on subjects at rest, without controlled thoughts, enable including impaired subjects. In such situations, blind mining algorithms are used to uncover intrinsic brain structure and to relate it to subjects' cognitive or clinical profile.
I'll discuss the computational statistics models as well as the successes in characterizing individuals' neuropsychiatrics.
This research program leads to a big-picture methodological question: how can we assemble these data-mining tools to refine representations of cognition from brain data. Beyond cognitive neuroscience, data-intensive investigations open the door to quantitative research in new fields. But it requires a shift in scientific methodology: redefining the accepted notion of a model, reinventing validation without experimental intervention, linking data-driven findings to non-formalized domain knowledge.
INRIA Parietal
NeuroSpin/CEA Saclay