BEGIN:VCALENDAR
PRODID:-//TAMU Math Calendar//NONSGML v1.0//EN
VERSION:2.0
BEGIN:VEVENT
DTSTART:20240213T160000Z
DTEND:20240213T170000Z
SUMMARY:Fast Tensor Methods and Implicit Computations
DESCRIPTION:In this talk I will focus on fast methods for analysing and decomposing tensor data. In the first part of the talk\, I will introduce a method we proposed for symmetric tensor decomposition. We provide several guarantees for the algorithm and its relevant non-convex optimization problem. Furthermore\, we observe empirically that the method is roughly one order of magnitude faster than existing decomposition algorithms\, and is also robust to noise. On the second part of the talk\, I will cover implicit method of moments. Higher-order moments of multivariate random variables suffer from a curse of dimensionality: the number of entries scale exponentially with the order of the moments. We introduce an implicit approach that allows for estimating parameters without explicitly forming the moments\, that way avoiding the curse of dimensionality. We use this approach to estimate the parameters of Gaussian Mixture Models\, obtaining a method with computational and storage costs similar to those of state-of-the arts methods\, such as expectation-maximization\, and opening the door to the practical use of the method of moments for multivariate variables. Finally\, I will mention several related methods and applications\, including on-going work on using the method introduced in the first part of the talk for decomposing moment tensors implicitly.
END:VEVENT
END:VCALENDAR