Machine Learning for Neuroimaging Talk, by Bertrand Thirion, NeuroSpin, Paris, Thursday April 24, 10:00-11:00, at DTU, Building 321, room 033.
Abstract
Accumulating the information conveyed by brain images, such as functional Magnetic Resonance Images (fMRI), to build brain templates, emains one of the major challenges faced by neuroimaging. Due to the difficulty of sharing, gathering and analyzing large sets of images, coordinate-based meta-analyzes, that rely on published brain locations, have become the de facto standard to interrogate existing datasets. While being very useful, this exploration is limited by the large amount of information dropped by and the abstraction inherent in the coordinate-based representation. By contrast, recent image-based approaches, that build on the increase of publicly available data, are expected to give a more reliable picture of brain functional organization. In this talk, we will explore two methodological aspects of image-based brain meta-analyzes: i) the estimation of spatial components from large sets of images, based on sparse online principal components analysis techniques ii) the reverse inference problem, where some information is tentatively inferred from brain activation patterns, which results in a hard multi-label learning problem.