Learning with Low-rank Approximations
Title : Learning with Low-rank Approximations
Asbtract : Matrix and tensor factorizations are widespread techniques to extract structure out of data in a potentially blind manner. Research on tensor methods is rapidly growing and encompases many aspect of computer science such as high performance computing and large scale non-convex optimization. An important challenge is to propose or study matrix and tensor models which are of practical interest while making efficient use of recent developements in both low-level tensor computation techniques such as tensor contractions on GPUs and large-scale non-convex optimization techniques such as stochastic gradient algorithms and proximal algorithms. After introducing low-rank approximation methods and depicting the current research landscape, I will focus on two recent contributions: (i) Unsupervised music automatic segmentation using nonnegative Tucker decomposition (ii) Heuristic extrapolated block-coordinate descent algorithm for tensor decompositions.
Contact : https://jeremy-e-cohen.jimdofree.com/
Séminaire en ligne : https://lpensl.my.webex.com/
Room : webconf3.physique