Low Rank Tensor Recovery via Cubic Sketching

Date
10/27/2017
Location
11:30AM at LWSN B134
Speaker
Prof. Guang Cheng, Purdue University
Title
Low Rank Tensor Recovery via Cubic Sketching
Host
Description

In this presentation, we propose a general framework for recovering sparse and low-rank tensors through rank-one cubic sketching. Two real-world applications are considered: one on high-dimensional interaction models; another on compressed image transmission. A block-wise thresholded gradient decent algorithm is proposed for stable recovery in both noiseless and noisy cases. Both upper bound and lower bounds for the estimation accuracy are obtained over a large class of low-rank tensors, demonstrating the optimality of the proposed procedure. This is an ongoing work with Botao Hao and Anru Zhang.