Goal of PARTENSOR is twofold:
1) Development of a Parallel Toolbox for processing very large tensors. Recently, a significant research activity has been developed at the Tenchnical University of Crete in the area of Tensor Factorization (TF). This activity has led to 1) publications at leading scientific conferences and journals and 2) parallel implementations through C++ and Message Passing Interface (MPI), using TF algorithms in distributed memory systems. In addition to TF, the toolbox will include a list of factorization and completion algorithms, such as Tucker, INDSCAL, PARAFAC2, depending on the nature of the data. Furthermore, it will be possible to impose structure on latent factors, such as sparsity, symmetry, orthogonality, etc.
To maximize the applicability of the toolbox to multi-processor systems of various kinds (distributed or shared memory, CPU-GPU combination, etc.), we will develop implementations of the algorithms in MPI, OpenMP, and OpenCL.
2) Development of parallel tensor algorithms for fMRI processing. Functional Magnetic Resonance Imaging (fMRI) is the most modern and popular method of mapping the human brain with important applications in clinical practice (neurology and neurosurgery). Analysis of fMRI data is extremely demanding due to the combination of large volumes of data and low signal-to-noise ratio. The use of tensor models for fMRI data analysis has attracted considerable research interest recently because it retains its multidimensional structure. However, in general, classical tensor models are not fully compatible with fMRI data. Our goal is to construct tensor models suitable for modeling fMRI data, and to develop parallel algorithms for tensor factorization or completion.
Within PARTENSOR, an assessment of the validity of the models to be developed in real data from a sufficient number of activation tests (motor and mental) and individuals is provided, as compared to existing, standard processes for the analysis and export of activation maps.