Microkicking for Fast Convergence of Sparse Kaczmarz and Sparse LMS
Sprache des Vortragstitels:
2017 IEEE International Workshop on Computational Advances in Multi-Sensor Adaptive Processing
Sprache des Tagungstitel:
Algorithms based on linearized Bregman iterations are able to perform sparse reconstruction at a low computational complexity. Especially the Least-Mean-Squares (LMS) and Kaczmarz variants of linearized Bregman iterations proved to be very feasible for fixed-point digital hardware implementation. We present a method that we call microkicking for improving the convergence speed of linearized Bregman based algorithms. This method can be implemented with only a negligible complexity overhead leading to significantly faster convergence for both variants of the linearized Bregman iterations. We furthermore show simulation results demonstrating
the performance gains achievable by microkicking.