SoTT: a greedy construction of a sum of Tensor Trains - Inria - Institut national de recherche en sciences et technologies du numérique Accéder directement au contenu
Communication Dans Un Congrès Année : 2023

SoTT: a greedy construction of a sum of Tensor Trains

Résumé

Tensor Train is probably the tensor format most widely used in a number of fields, ranging from physics to machine learning. In this work, we focus on the compression problem of a given multi-way datum. It is known that in Tensor Train the order of the variables to be treated has to be fixed a priori, and this might have a non-negligible influence on the compression performances. We propose a compression method in which a given datum is compressed as a sum of Tensor Trains. Each term of a sum is a Tensor Train in which we do not fix a priori neither the order of variables nor the ranks: these are determined based on error and parsimony criteria, taken into account through the introduction of a trade-off function. We prove that, for finite dimensional data, the trade-off function chosen makes it possible to certify the approximation (in terms of accuracy) and get a precise convergence rate. Several encouraging numerical results were proposed on moderate order tensors (up to 6th order).
Fichier non déposé

Dates et versions

hal-04392369 , version 1 (13-01-2024)

Licence

Paternité

Identifiants

  • HAL Id : hal-04392369 , version 1

Citer

Damiano Lombardi, Virginie Ehrlacher, Maria Fuente-Ruiz. SoTT: a greedy construction of a sum of Tensor Trains. ICOSAHOM 2023 - International Conference on Spectral and High Order Methods, Yonsei University, Seoul, Aug 2023, Seoul, South Korea. ⟨hal-04392369⟩
27 Consultations
0 Téléchargements

Partager

Gmail Facebook X LinkedIn More