Hosted by: Kerstin Kleese van Dam

Matrix and tensor factorization techniques are important data analysis tools with numerous applications in recommender systems, text processing, data mining, and image processing. Matrix factorization is a dimension reduction method that extracts features and provides low-rank matrix approximation; while tensor factorization extends these properties to multidimensional arrays. In this talk, we will discuss the following three topics: 1) An efficient algorithm that improves time efficiency and scalability of tensor factorization. We will present a technique to speed up alternating least squares and gradient descent - two commonly used strategies for tensor factorization. By using the properties of Khatri-Rao product, we show how to efficiently address a computationally challenging sub-step of both algorithms, and how to implement the algorithm on parallel machines. 2) Application of matrix factorization for hand pose estimation. We will discuss how a joint matrix factorization and completion algorithm can be used to estimate the unknown joint angle parameters in hand pose estimation. We will conclude the talk by discussing future work on matrix factorization in streaming mode with limited memory. 3) Asynchronous matrix factorization. We will discuss how the distributed matrix factorization avoids bulk synchronization after every iteration.