• brain
• cognitive neuroscience
• fMRI
• multivariate analysis

Historically, our understanding of the human brain has been mutually affected both by the available methodologies and by the models of its functioning. Early neuroimaging studies, based on Magnetic Resonance Imaging (MRI), Positron Emission Tomography (PET) or Electroencephalography (EEG), tried to isolate specific cognitive operations and focused on subtractive models following a modular account of brain organization, in which individual, localized regions subserved distinct functions.
Nowadays, growing evidence support a much more complex view of the human brain as a dynamic network with multiple and integrated levels of organization. Multivariate machine learning techniques were developed to model complex, sparse neuronal populations and to push the explanatory power of neuroimaging far beyond the capabilities of classical inference, thus positively affecting cognitive neuroscience.
This thesis introduces a description with advantages and limitations of the most recent multivariate techniques and then presents examples of these novel approaches to functional MRI (fMRI) data of three different brain functional studies. Specifically, decoding, encoding and representational similarity techniques are described and applied to experimental brain functional data during visual perception of different classes of actions, motor execution of a large set of grasping gestures and conceptual representations of nouns. Finally, possible translational applications for clinical purposes are described both for characterizing neuropsychiatric disorders, pain prediction and other pathological conditions.