Acceleration of Optimization-based Structured Sparse Time-Frequency Analysis by ADMM

Keidai ARAI, Kohei YATABE

Session: General

Keywords: structured sparsity, convex optimization, perspective function, Nesterov acceleration, optimal step-size

Abstract: Sparse time-frequency (T-F) analysis has been studied to obtain localized T-F representations of a signal. Among various methods, optimization-based methods (e.g., basis pursuit) offer flexibility in designing T-F representations by designing the objective function. In particular, the convex-optimization-based method that imposes a specified structure on the magnitude of a T-F representation has realized T-F analysis tailored for specific applications. However, the conventional method uses the PDS (primal-dual splitting) algorithm, which is known to require many iterations in some cases. In this paper, we propose applying ADMM (alternating direction method of multipliers) with acceleration techniques to reduce the number of iterations required for obtaining structured sparse T-F representations. Experiments show that the proposed algorithm can obtain the T-F representation much faster than the conventional PDS algorithm.

Submission Number: 82