Título: Compressing Tensors for the Canonical Polyadic Decomposition
Palestrante: Felipe Diniz (IM-UFRJ)

Data: 11/09/2019 (quarta-feira)
Horário: 10:10
Local: IM-UFRJ, CT, sala C-116

Resumo: Given a tensor T and a rank R, the computation of a rank-R CPD is a task which we know to be quite costly. In order to mitigate these costs, one idea is to pre-process T, obtaining a smaller tensor, S, compute the CPD for this tensor, and from this CPD we can quickly recover a CPD for T. A common approach to this problem should be based on some kind of dimensionality reduction for tensors, and this is possible with the MLSVD (multilinear singular value decomposition). This is even the first step taken by Tensorlab before computing a CPD. Curiously, no other solver does this kind of dimensionality reduction. They just compute a CPD for the "raw" tensor, which can be very costly.