Project Overview

Our clients, Professors Rebecca Willett (Electrical and Computer Engineering), Robert Nowak (Electrical and Computer Engineering), and Timothy Rogers (Psychology), have requested assistance in conducting an interdisciplinary neural decoding project. In it, they are working to understand how modeling assumptions impact multivariate fMRI analysis. In the neuroimaging literature, there exists a high degree of variability in how multivariate analyses are conducted. With a vast array of machine learning tools to choose from and few direct comparisons between them, there does not exist a consensus on how or when to use a particular method.

To understand why this is a significant issue, consider a researcher trying to determine whether the brain’s response to a stimulus is anatomically localized or distributed. Many popular analysis methods implicitly or explicitly assume localized responses. Without understanding the tradeoffs underlying different analysis methods, we are ill-equipped to assess the validity of these assumptions, illustrating a fundamental gap in our understanding of the mechanisms of neural functioning.

Our clients seek to address this problem by benchmarking the performance of many multivariate pattern classifiers across a range of publicly available datasets. Evaluating the predictive power of different algorithms provides a means to empirically test hypotheses of neural activity implicit to different modeling assumptions. However, this study would be infeasible if conducted on standard computing resources, as a single analysis has been reported to take as much as 15 years of computing time.

We are working with Professors Willett, Nowak, and Rogers to develop a suite of software tools to conduct the aforementioned analyses with distributed computing resources. We propose a design which takes advantage of the HTCondor job scheduler to leverage enormous amounts resources made available through the Center for High Throughput Computing(CHTC) and Open Science Grid(OSG). Additionally, our design will overcome organizational challenges through use of a web-application which will serve to track data provenance and facilitate the use of CHTC and OSG computing resources. Finally, the design will be centered around a document based database, MongoDB, which will serve as an easily accessible repository to collect the results from completed analyses.

Team Picture

Contact Information

Team Members

• Ian Kinsella - Team Leader
• Elliott Janssen Saldivar - Communicator
• Laura Xu - BSAC
• Alison Walter - BWIG
• Zachary Petersen - BPAG

Advisor and Client

• Mitchell Tyler - Advisor
• Dr. Rebecca Willett - Client
• Chris Cox - Alternate Contact

Related Projects

• Spring 2016: Neural decoding software suite
• Fall 2015: Neural decoding software suite