Koby Hayashi will be working to compute constrained low-rank approximations of large-multidimensional fMRI brain scan data-sets. In particular, he will be optimizing and applying the Coupled Matrix Tensor Factorization (CMTF) algorithm. This work is with Professor Ballard and Dr. Michael Tobia. Funding Source: National Science Foundation.
Bel LaPointe and Edward Zhao: The Graduate School of Arts and Sciences has awarded Summer Research Fellowships to Bel LaPointe and Edward Zhao. Bel is looking to optimize a CHC genetic algorithm that creates gene interaction models from experimentally measured gene transcripts. Bel is working with Professors John, Norris and Muday. Edward is working with fNIR data collected from the prefrontal cortex and a rather traditional genetic algorithm to create an optode interaction model. Edward is working with Professors John, Norris, Rejeski and Laurienti.
Sajant Anand received an honorable mention for a Goldwater Scholarship. Also, he has been accepted to Harvard University’s SEAS NSF REU Summer Research Program this summer.
Steven Ma was awarded a Wake Forest Summer Research Fellowship with mentor, Dr. Todd Torgersen.
Yujie (Jeffrey) Jiang
Funding Source: Atlantic Coast Conference Inter-institutional Academic Collaborative (ACCIAC) Fellows Program
Abstract: Today’s problems in machine learning and applications such as signal processing often involve massive amounts of multi-dimensional data. The goal of this project is to develop efficient algorithms for analyzing these multi-dimensional data sets. He will focus on the computational kernels within tensor factorization, which is a fundamental form of unsupervised machine learning that decomposes multi-dimensional data into a low-rank representation in order to apply subsequent analysis. On today’s computers, the cost of data movement (e.g., how well algorithms exploit the cache hierarchy) is becoming a determining factor for the speed of computation. The approach of the project is to reformulate current algorithms in order to move less data and run faster, allowing for interactive data analytic tools and the ability to tackle larger data sets.
Irina Viviano
Funding Source: WFU Scholars Program (Reynolds Scholarship)
Abstract: In applications such as computer vision and neuroscience, data often manifests itself in the form of multi-dimensional arrays known as tensors. In order to provide meaningful insights and analysis into the data, tensor decompositions are often used. One such decomposition is the CANDECOMP/PARAFAC (CP) decomposition, which decomposes a tensor into a sum of component rank one tensors that are more easily digestible. In practice, the CP decomposition is computed via the alternating least squares (ALS) method, reducing the problem to several linear least squares problems, which are then solved using the normal equations. However, the normal equations method is susceptible to numerical ill-conditioning, which could compromise the accuracy of the results and the conclusions of the analysis. The goal of this project is to explore the feasibility of exploiting the QR-factorization, a more numerically stable alternative to the normal equations, in order to improve the accuracy of the ALS algorithm without sacrificing the computational time required to solve the problem.
