Tensor train (TT) decomposition provides a space-efficient representation for
higher-order tensors. Despite its advantage, we face two crucial limitations
when we apply the TT decomposition to machine learning problems: the lack of
statistical theory and of scalable algorithms. In this