[Swift-commit] r2406 - text/hpdc09submission

[noreply at svn.ci.uchicago.edu]

Author: benc
Date: 2009-01-09 02:06:03 -0600 (Fri, 09 Jan 2009)
New Revision: 2406

Added:
   text/hpdc09submission/omxFigure.jpg
Modified:
   text/hpdc09submission/paper.latex
Log:
import skenny's summary of SEM app

Added: text/hpdc09submission/omxFigure.jpg
===================================================================
(Binary files differ)


Property changes on: text/hpdc09submission/omxFigure.jpg
___________________________________________________________________
Name: svn:mime-type
   + application/octet-stream

Modified: text/hpdc09submission/paper.latex
===================================================================
--- text/hpdc09submission/paper.latex	2009-01-08 21:12:11 UTC (rev 2405)
+++ text/hpdc09submission/paper.latex	2009-01-09 08:06:03 UTC (rev 2406)
@@ -1027,6 +1027,63 @@
 it extracts the file name of input header and output header, which are
 then put in the command line to invoke the reorient program.
 
+\subsection{Structural Equation Modeling using OpenMx}
+
+OpenMx is an R library designed for structural equation modeling (SEM),
+a technique currently used in the neuroimaging field to examine
+connectivity between brain areas.
+
+Structural Equation Modeling is greatly enhanced when coupled with
+Grid-resources and a workflow management system. Traditionally,
+structural equation models have been derived from anatomical models
+based on primate brains out of necessity. It was infeasible to test
+models outside of the hypothetical anatomical model space due to
+restrictions in resources. In the current infrastructure we implement a
+scriptable, high-level means for not only testing but generating
+exploratory models in parallel on large clusters, making the testing of
+models outside the (anatomical) hypothesis space a more reasonable goal.
+In light of this, we have developed a ``model generator'' to allow a
+researcher to test all models within a space of potential connections
+without a predefined anatomical model. In the absence of a large-scale
+infrastructure this would not be doable. For example, within the CNARI
+submit cluster at Chicago, we have a relatively simple Swift script for
+calling OpenMx to generate and process models in parallel.
+
+\begin{figure}[htbp]
+\includegraphics{omxFigure}
+
+\caption{Schematic of a single OpenMx model containing 4 regions of
+interest (I through K) with 5 regression starting values (asymmetric
+connections) of weight 0.75 and 4 residual variances (symmetric connections)
+of weight 1.0}
+\end{figure}
+
+Using OpenMx's model generator -- a set of functions which creates
+self-contained, structural equation models -- we generated 65,535 R
+objects representing all models with anywhere from 1 to 16 connections
+of varying weights between 4 pre-selected regions of interest. That is,
+a 4x4 matrix represents connections between the four regions,
+representing 16 possible connections (connections between the same two
+regions but in different directions are tested separately). We queried
+our experiment database for activation values based on the selected
+regions of interest, and the covariance of those regions over 8 time
+points during the emblem experiment was calculated. The covariance of
+each generated model was then compared to the covariance matrix of the
+observed data to determine the best-fitting model. In other words, the
+connection weights (or strength of the relationships between anatomical
+regions) can be explored based on the fit of each model.
+
+modgenproc.swift is used to submit each of the necessary computation
+components to TeraGrid's Ranger cluster: a) the model object b) the
+covariance matrix derived from the database and c) the R script which
+makes the call to OpenMx. Once the job is assigned to a node, OpenMx?
+estimates weight parameters for each connection within the given model
+that results in a model covariance closest to the observed covariance of
+the data. Each of these compute jobs returns its solution model object
+as well as a file containing the minimum value achieved from that model.
+The processing of these models on Ranger was achieved in <45 minutes.
+
+
 \section{Usage Experience}
 
 \subsection{Use on large numbers of sites in the Open Science Grid}