Author:

Keywords:

PSI_MBL, PSI_VISICS, PSI_4307

Abstract:

The goal of the thesis is to propose methods for learning sparse and structured models from data that can be naturally represented as a graph. Our techniques are targeted towards the setting of large dimensional data and few samples. In particular we apply many of our methods to the problem of understanding and interperting high dimensional functional MRI data. To this end we propose a) A method for learning an interpertable predictive linear model with graph based constraints b) a novel approach for discovering the graph of interations between high dimensional data using an easy to define prior c) a principled approach to discover the difference of interactions in similar data and characterize the uncertainty of the finding. We apply our methods to real neuroimaging datasets, natural image datasets, as well as genetics data. Finally we explore the application of graph structured prediction methods in the setting of natural image retrieval.