FoCM 2014 conference
Workshop C3 - Learning Theory
December 19, 15:00 ~ 15:30 - Room B23
Tensor decomposition, convex optimization, and multitask learning
Ryota Tomioka
Toyota Technological Institute at Chicago, USA - tomioka@ttic.edu
Tensor factorization, or multilinear modelling, has received much attention recently. Compared to its two-dimensional counterpart, matrix factorization, many properties related to tensors, for example, the rank, are known to be hard to compute. Recently new approaches based on convex relaxation of tensor (multilinear-)rank have emerged. Although, these new methods come with worst case performance guarantees, they tend to be less efficient than previously known greedy algorithms in practice. I will overview and discuss the possibility and limitation of these approaches from the perspective of computation-statistics trade-off. Furthermore, I will present a recent application of the above idea to multi-task learning.
Joint work with Kishan Wimalawarne (Tokyo Institute of Technology,), Taiji Suzuki (Tokyo Institute of Technology), Kohei Hayashi (National Institute of Informatics, Tokyo), Hisashi Kashima (Kyoto University) and Masashi Sugiyama (University of Tokyo).