News Archives

RSS Feed

  • UNM
  • >Home
  • >News
  • >2006
  • >February
  • >Discrete Bayesian Network Structure Search with an Application to Functional Magnetic Resonance Imaging Data

Discrete Bayesian Network Structure Search with an Application to Functional Magnetic Resonance Imaging Data

February 20, 2006

  • Date: Tuesday, February 20, 2006 
  • Time: 11:00 am — 12:15 pm
  • Place: Woodward 149

John Burge
Department of Computer Science, UNM

Neuroimaging technologies, such as functional magnetic resonance imaging (fMRI), have become widely used in the analysis of mental illness. They are non-invasive techniques used to create images that measure (possibly indirectly) human neural activity. Measuring something as complex and highly detailed as the human brain poses significant data mining challenges. For example, an fMRI scan for a single patient can posses extremely high dimensionality, easily resulting in more than 250 megabytes of data. Neuroscientists and statisticians have a wide range of methods for analyzing such data, however, many of these methods make linear assumptions that are not likely true given the dynamics of the human brain. We introduce a nonlinear method for modeling the data based on discrete Bayesian networks, a modeling framework commonly employed within the machine learning community. We provide a synopsis for this framework and discuss the extensions we have proposed for learning generative and class-discriminative models as well as for improving model selection based on the hierarchical arrangement of neuroanatomical regions of interest. We briefly discuss our experimental results and methods used to validate them.