Structured-joint factor estimation for high-order and large-scale tensors

Overview & Synopsis

Multidimensional data sets and signals occupy an important place in recent application fields. Tensor decomposition represents a powerful mathematical tool for modeling multidimensional data and signals, without losing the interdimensional relations. The Canonical Polyadic (CP) model, a widely used tensor decomposition model, is unique up to scale and permutation indeterminacies. This property facilitates the physical interpretation, which has led the integration of the CP model in various contexts. The main challenge facing the tensor modeling is the computational complexity and memory requirements. High-order tensors represent a important issue, since the computational complexity and the required memory space increase exponentially with respect to the order. Another issue is the size of the tensor in the case of large-scale problems, which adds another burden to the complexity and memory. Tensor Networks (TN) theory is a promising framework, allowing to reduce high-order problems into a set of lower order problems. In particular, the Tensor-Train (TT) model, one of the TN models, is an interesting ground for dimensionality reduction. However, respresenting a CP tensor using a TT model, is extremely expensive in the case of large-scale tensors, since it requires full matricization of the tensor, which may exceed the memory capacity.In this thesis, we study the dimensionality reduction in the context of sparse-coding and high-order coupled tensor decomposition. Based on the results of Joint dImensionality Reduction And Factor rEtrieval (JIRAFE) scheme, we use the flexibility of the TT model to integrate the physical driven constraints and the prior knowledge on the factors, with the aim to reduce the computation time. For large-scale problems, we propose a scheme allowing to parallelize and randomize the different steps, i.e., the dimensionality reduction and the factor estimation. We also propose a grid-based strategy, allowing a full parallel processing for the case of very large scales and dynamic tensor decomposition.