Adaptive Step Size Momentum Method For Deconvolution

Abstract

In this paper, we introduce an adaptive step size schedule that can significantly improve the convergence rate of momentum method for deconvolution applications. We provide analysis to show that the proposed method can asymptotically recover the optimal rate of convergence for first-order gradient methods applied to minimize smooth convex functions. In a convolution setting, we demonstrate that our adaptive schedule can be implemented efficiently without adding computational complexity to traditional gradient schemes.