Here we report our approach to the “Pittsburgh Brain Activity Interpretation Competition 2.pdf

Pittsburgh Brain Activity Interpretation Competition 2006 Methods Description Brain Reading with Selective Training Data, Transduction, and Behavior Based Group Analysis Project Abstract Here we report our approach to the “Pittsburgh Brain Activity Interpretation Competition 2006” in which the goal was to predict the responses of three subjects who were watching video scenes during a functional magnetic resonance imaging (fMRI) study. Our work uses the continuous valued output from support vector machine classification (SVC) models to estimate a heterogeneous set of temporal “features” as rated by three subjects in response to audio/video stimuli. The primary challenge to this approach was to provide the SVC with appropriate class labels for each feature. As part of our approach, we attempted to fully utilize the unique properties of the data and the parameters of the competition, which included the existence of interdependencies between the feature ratings, a high degree of consistency in the ratings across subjects, and access to the target data (season 3) images. Primary conclusions from our work indicate that 1) the quality of the labels used for training the SVCs is more important than both the quantity of training data and image data variable reduction/selection, and 2) human behavior (in this case feature rating responses to multimedia stimuli) can be much more consistent across individuals than brain structure or cognitive strategy and the behavior’s temporal nature makes it inherently much easier to “align” than spatial brain data. We are excited by the potential of this work to provide a new approach to conceptualizing group analyses in the field of neuroimaging. Introduction Determining relationships that allow us to generalize beyond our direct experience is the primary goal of scientific exploration as well as a critical aspect of daily life. Often it is the availability and quality of observational data that direct