Skip navigation
Please use this identifier to cite or link to this item:
Full metadata record
DC FieldValueLanguage
dc.contributor.advisorRamadge, Peter Jen_US
dc.contributor.authorXu, Haoen_US
dc.contributor.otherElectrical Engineering Departmenten_US
dc.description.abstractMachine learning is the science of discovering knowledge from data and making decisions based on this learnt structure. It has grown into a broad discipline that has established many statistical and computational learning principles. We are surrounded by the applications of machine learning every day: speech recognition, recommendation systems and many more. This thesis addresses the application of machine learning to the discovery of hidden patterns in functional magnetic resonance imaging (fMRI) data and developing practical algorithms exploiting these patterns. Over the years, machine learning has been introduced to many data-intensive empirical sciences. Since the success of multi-voxel pattern analysis (MVPA) in 2001, machine learning has been steadily making its way into fMRI data analysis field. fMRI is a non-invasive technique to indirectly investigate brain activity. It records brain activity as a time sequence of 3-D images such that the exhibited patterns under different tasks can be visualized and studied. Since fMRI data is noisy and high-dimensional, the ability to extract useful information effectively and efficiently is critical. We start with the application of multi-set fMRI data alignment. This is an essential step in identifying key characteristics representative of multiple subjects. A previously-proposed method, hyperalignment, has shown significant improvement over traditional anatomical alignment. We introduce its regularized extension and connect hyperalignment with canonical correlation analysis. This further improves alignment results. Then we look into the generalized lasso. It has potential applications in spatially informed analysis of fMRI data. We demonstrate that the generalized lasso problem is reducible to a subspace constrained lasso. We also show that it can be reduced to a standard lasso in the dual space by dictionary filtering. Lastly, we present a real-time learning system for fMRI data. It carries out classification and prediction of brain states simultaneously with fMRI data acquisition. This is beneficial to understanding how the brain processes information, building an interactive experimental paradigm and many new aspects in fMRI study. The core to this real-time fMRI system is an online conjugate gradient algorithm. It can process high-dimensional fMRI data efficiently and reach classification accuracy comparable to traditional off-line learning methods.en_US
dc.publisherPrinceton, NJ : Princeton Universityen_US
dc.relation.isformatofThe Mudd Manuscript Library retains one bound copy of each dissertation. Search for these copies in the <a href=> library's main catalog </a>en_US
dc.subjectfMRI data analysisen_US
dc.subjectfunctional alignment of multi-set fMRI dataen_US
dc.subjectmachine learningen_US
dc.subjectreal-time fMRI learningen_US
dc.subjectspatially informed fMRI analysisen_US
dc.subject.classificationElectrical engineeringen_US
dc.titleApplications of Machine Learning to FMRI Data Analysisen_US
dc.typeAcademic dissertations (Ph.D.)en_US
Appears in Collections:Electrical Engineering

Files in This Item:
File Description SizeFormat 
Xu_princeton_0181D_10676.pdf4.09 MBAdobe PDFView/Download

Items in Dataspace are protected by copyright, with all rights reserved, unless otherwise indicated.