Melissa Carroll will present her preFPO at 2PM on Monday October 12 in Room 402. The members of her committee are: Rob Schapire, advisor; Ken Norman (Psychology) and Irina Rish (IBM), readers; David Blei and Peter Ramadge (ELE), nonreaders. Everyone is invited to attend her talk. Her abstract follows below. --------------------- Capturing Spatial Structure of Response in Predictive fMRI Models Functional MRI (fMRI) is commonly used to map between brain regions and mental functions. Traditional analyses have focused on testing putative associations between large, known brain regions and broad mental tasks; however, machine learning techniques applied globally at the scale of 3D voxels have been used to produce fine-scaled predictive associations. Extracting meaningful insight into brain structure-function relationships from these techniques remains a challenge, though, partially because voxels are arbitrary discretizations of the underlying activation signal, leading to spatial auto-correlation and interfering with model interpretation. I discuss methods that address this spatial structure with the goal of achieving well-predicting models that are also reliable and easily interpreted. I first demonstrate that sparse regression methods produce more accurate mental state predictions than non-sparse methods. Furthermore, I show that the Elastic Net hybrid regularized regression penalization produces models that select more reliable sets of voxels across experimental runs than the standard LASSO, without compromising prediction performance. This increased reliability comes in part from greater inclusion from within spatially localized clusters of activity, and the most predictive models found are those exhibiting distributed patterns of localized activity. Given the local clustering properties of the fMRI signal, a natural way to improve reliability further might be to directly find these clusters. Spatial filters have traditionally been used to smooth fMRI data and reduce noise; however, the optimal filter varies by task, region, and subject, and training models to find the most predictive filters can be computationally prohibitive. I describe a parallel implementation of the LARS-EN algorithm that can exploit massively parallel architectures, and demonstrate the feasibility of using it to build models from hundreds or thousands of spatial filters, e.g. Gaussian or wavelet. I show how this framework can be used as a test-bed for evaluating filter sets, with the potential to produce more reliable models without compromising prediction performance or parsimony. I also show how these models can be interpreted to elucidate spatial properties of the neural response. Finally, I introduce an alternative method that can potentially model more precise Gaussian spatial filters within a flexible framework. This approach directly determines filter parameters by treating the fMRI activity as a distribution in space, characterized as a Mixture of Gaussians. Unlike recent similar models, I show that by assuming a Bayesian hierarchical model for the data similar to that of Latent Dirichlet Allocation, one can exploit developments in variational inference algorithms for supervised LDA to tractably infer such probabilistic models within the desired predictive modeling framework.