Speaker: Cagri Ozdemir, Ph.D. candidate in Data Science and Engineering, Department of Electrical Engineering and Computer Science, South Dakota School of Mines and Technology

 

Title: Advancing Machine Learning Through Multilinear Subspace Methods

Abstract: The last few decades have seen significant advances in various sensing instruments generating massive volumes of n-dimensional data. To best explore, analyze, and provide insights from such data, new mathematical tools are needed in an effort to bridge the gap between traditional machine/deep learning

models and their multilinear counterparts. This talk presents a new approach to dealing with such data using a multilinear (tensor-tensor) perspective and provides insight into how such a bridge can be built. In particular, we formulate the mathematical theory of tensor decomposition via an algebra of circulants detailing novel extensions of traditional linear algebraic tools. We then provide insights into several different application ares within the machine/deep learning community and illustrate how multilinear extensions can be achieved. We conclude with a discussion surrounding such developments and possible application spaces not yet investigated.