A Momentum Block-Randmoized Stochastic Algorithm for Low-Rank Tensor CP Decomposition

发布时间：2021-04-15 作者：浏览次数：

Speaker:	崔春风	DateTime:	2021年4月23号（周五）下午2:00-3:00
Brief Introduction to Speaker: 崔春风，北京航空航天大学教授。		Place:	腾讯会议（会议号请联系张雄军老师索取）
Abstract:The block-randomized stochastic algorithm has shown its power in handling high-dimensional low-tank tensor canonical polyadic decomposition (CPD). Since computing CPD is computationally expensive, there is great interest in speeding up the convergence. In this talk, we introduce a momentum accelerated version of the block-randomized stochastic gradient descent (SGD) algorithm for low-rank tensor CPD. Under some mild conditions, we show the global convergence to the stationary point of the fixed stepsize algorithm for this nonconvex nonsmooth optimization problem. Compared with the algorithms without momentum, the preliminary numerical experiments for the synthetic and real data demonstrated that our accelerated algorithms are efficient, and can achieve better performance in terms of objection function value, mean squared error, and structural similarity value. This talk is based on the joint work with Qingsong Wang and Deren Han.