Multilinear Subspace Learning

This project aims to host multilinear subspace learning (MSL) algorithms for dimensionality reduction of multidimensional data through learning a low-dimensional subspace from tensorial representation directly.

The origin of MSL traces back to multi-way analysis in the 1960s and they have been studied extensively in face and gait recognition. With more connections revealed and analogies drawn between multilinear algorithms and their linear counterparts, MSL has become an exciting area to explore for applications involving large-scale multidimensional (tensorial) data as well as a challenging problem for machine learning researchers to tackle.

MSL-based dimensionality reduction employs either tensor-to-tensor projection (TTP) or tensor-to-vector projection (TVP). TTP finds a direct mapping from high-dimensional tensors to low-dimensional tensors of the same (or lower) order. TVP finds a direct mapping from high-dimensional tensors to low-dimensional vectors.