EVENT DETAILS AND ABSTRACT

Title: Modewise tensor dimension reduction and applications
Speaker: Elizaveta Rebrova
Speaker Info: Lawrence Berkeley National Lab
Brief Description:
Special Note:
Abstract:

Although tensors are a natural multi-modal extension of matrices, going beyond two modes (that is, rows and columns) presents many interesting non-trivialities. For example, the notion of singular values is no longer well-defined, there are various competing versions of the tensor rank, and while it is tempting to work with the tensor data using known low-modal methods (just "unwrapping" a given tensor to a very long vector or rearranging it to a matrix), this approach is inefficient in several ways. The main goal of my talk is to make these statements precise and to present and discuss modewise randomized projection methods, preserving both structure and geometry of a tensor while reducing its size. Then, I will talk about several applications when this turns out to be very useful, including tensor low-rank fitting, compressive sensing and topic modeling tasks.