This uses a sub-string search, so that "graph" matches both "graph" and "graphical".
This search only searches by keyword. If you'd like, you can also search by researcher.
Found 5 results.
Listing from newest to oldest
TR-CS-2010-08
Exploiting Task Relatedness to Learn Multiple Bayesian Network Structures
Diane Oyen and Terran Lane
We address the problem of learning multiple Bayesian network structures for experimental data where the experimental conditions define relationships among datasets. A metric of the relatedness of datasets, or tasks, can be described which contains valuable information that we exploit to learn more robust structures for each task. We represent the task-relatedness with an undirected graph. Our method uses a regularization framework over this task-relatedness graph to learn Bayesian network structures for each task that are smoothed toward the structures of related tasks. Experiments on synthetic data and real fMRI experiment data show that this method learns structures which are close to ground truth, when available, and which generalize to holdout data better than an existing multitask learning method, learning networks independently, and learning one global network for all tasks.
TR-CS-2009-03
Approaches to the Network-Optimal Partition Problem for fMRI Data
Benjamin Yackley, Terran Lane
Although much MRI research has been performed using existing atlases of brain regions, it is possible that these regions, which are anatomically determined, are not the truly optimal way to divide brain activity into parts. The goal of this research is to discover a way to take existing fMRI data and use it as a basis for a clustering method that would divide the brain into functional partitions independent of anatomy. Since Bayesian networks over anatomical regions have been shown to be informative models of brain activity, this paper describes attempts to solve a mathematical problem which is an abstraction of the above goal: to partition a group of variables (the voxels of the MRI data) into clusters such that the optimal Bayesian network given the clustering is as "good" as possible.
TR-CS-2005-24
Dynamic Bayesian Networks: Class Discriminative Structure Search with
an Application to Functional Magnetic Resonance Imaging Data
John Burge
The neuroanatomical behavior and underlying root causes of many mental illnesses are only moderately understood. To further understand these illnesses, the relationships among neuroanatomical regions of interest (ROIs) can be modeled. Neuroscientists have several standardized methods which have yielded much success, however, many of these methods make restrictive assumptions. On the other hand, the machine learning community has developed techniques that do not require many of these assumptions. I propose a method based on dynamic Bayesian networks (DBNs) that is capable of modeling the temporal relationships among ROIs. DBNs are graphical models that explicitly represent statistical correlations among random variables (RVs). Within the fMRI domain, it is not known a priori which ROIs are correlated. Thus, the DBN's structure modeling the ROI interactions must be learned. My research focuses on novel techniques capable of learning the structure of DBNs from a given set of data. First, I observe that under certain circumstances, learning multiple Bayesian networks, one for each possible class of data, is a better approach than the frequently employed method of learning a single BN with a class node. Second, I propose a method for determining which structures are representative of class-discriminating relationships. Third, I propose a structure search heuristic that takes advantage of the relationships among hierarchically related RVs. Finally, I propose to incorporate shrinkage (a statistical technique previously used in machine learning) in the calculation of the DBN's parameters. This proposal also summarizes preliminary work on the classification of functional magnetic resonance imaging data with my proposed scoring function.
TR-CS-2004-28
Bayesian Classification of FMRI Data: Evidence for Altered Neural Networks in Dementia
John Burge, Vincent P. Clark, Terran Lane, Hamilton Link and Shibin Qiu
The alterations in functional relationships among brain regions associated with senile dementia are not well understood. We present a machine learning technique using dynamic Bayesian networks (DBNs) that extracts causal relationships from functional magnetic resonance imaging (fMRI) data. Based on these relationships, we build neural-anatomical networks that are used to classify patient data as belonging to healthy or demented subjects. Visual-motor reaction time task data from healthy young, healthy elderly, and demented elderly patients (Buckner et al. 2000) was obtained through the fMRI Data Center. To reduce the extremely large volume of data acquired and the high level of noise inherent in fMRI data, we averaged data over neuroanatomical regions of interest. The DBNs were able to correctly discriminate young vs. elderly subjects with 80% accuracy, and demented vs. healthy elderly subjects with 73% accuracy. In addition, the DBNs identified causal neural networks present in 93% of the healthy elderly studied. The classification efficacy of the DBN was similar to two other widely used machine learning classification techniques: support vector machines (SVMs) and Gaussian nave Bayesian networks (GNBNs), with the important advantage that the DBNs provides candidate neural anatomical networks associated with dementia. Networks found in demented but not healthy elderly patients included substantial involvement of the amygdala, which may be related to the anxiety and agitation associated with dementia. DBNs may ultimately provide a biomarker for dementia in its early stages, and may be helpful for the diagnosis and treatment of other CNS disorders.
TR-CS-2004-12
Parallel Kernel Computation for High Dimensional Data and Its Application to
fMRI Image Classification
Shibin Qiu and Terran Lane
Kernel method is frequently used for support vector machine classification and regression prediction. Kernel computation for high dimensional data demands heavy computing power. To shorten the computing time, we study the performance issues of parallel kernel computation. We design the parallel algorithms and apply them to the classification of fMRI images, which have extreme high dimensionality. We first formulate the kernel calculation through linear algebraic manipulation so that the operations can be performed in parallel with minimum communication cost. We then implement the algorithms on a cluster of computing workstations using MPI. Finally, we experiment with the fMRI data to study the speedups and communication behaviors.
News & Events | People | Academics | Research | Jobs | Companies | Visitors | Computer Facilities | Home
